Search Images Maps Play YouTube News Gmail Drive More »
Sign in
Screen reader users: click this link for accessible mode. Accessible mode has the same essential features but works better with your reader.

Patents

  1. Advanced Patent Search
Publication numberUS5640490 A
Publication typeGrant
Application numberUS 08/339,902
Publication dateJun 17, 1997
Filing dateNov 14, 1994
Priority dateNov 14, 1994
Fee statusLapsed
Also published asUS5873062
Publication number08339902, 339902, US 5640490 A, US 5640490A, US-A-5640490, US5640490 A, US5640490A
InventorsC. Hal Hansen, Dale Lynn Shepherd, Robert Brian Moncur
Original AssigneeFonix Corporation
Export CitationBiBTeX, EndNote, RefMan
External Links: USPTO, USPTO Assignment, Espacenet
User independent, real-time speech recognition system and method
US 5640490 A
Abstract
A system and method for identifying the phoneme sound types that are contained within an audio speech signal is disclosed. The system includes a microphone and associated conditioning circuitry, for receiving an audio speech signal and converting it to a representative electrical signal. The electrical signal is then sampled and converted to a digital audio signal with a digital-to-analog converter. The digital audio signal is input to a programmable digital sound processor, which digitally processes the sound so as to extract various time domain and frequency domain sound characteristics. These characteristics are input to a programmable host sound processor which compares the sound characteristics to standard sound data. Based on this comparison, the host sound processor identifies the specific phoneme sounds that are contained within the audio speech signal. The programmable host sound processor further includes linguistic processing program methods to convert the phoneme sounds into English words or other natural language words. These words are input to a host processor, which then utilizes the words as either data or commands.
Images(48)
Previous page
Next page
Claims(36)
What is claimed and desired to be secured by United States Patent is:
1. A sound recognition system for essentially real-time identification of, and in an essentially speaker independent manner, phoneme sound types that are contained within an audio speech signal, the sound recognition system comprising:
audio processor means for receiving an audio speech signal and for converting the audio speech signal into a representative audio electrical signal;
analog-to-digital converter means for digitizing the audio electrical signal at a predetermined sampling rate so as to produce a digitized audio signal; and
sound recognition means for identifying phoneme sound types contained within the audio speech signal, said sound recognition means comprising:
means for performing time domain analysis on a plurality of segmentized portions of the digitized audio signal so as to identify a plurality of time domain characteristics of the audio signal;
means for filtering each of the segmentized portions using a plurality of filter bands having predetermined high and low cutoff frequencies so as to identify thereby at least one frequency domain characteristic of each filtered segmentized portion; and
means for processing said time domain and frequency domain characteristics so as to identify therefrom the phonemes contained within the audio speech signal.
2. A sound recognition system as defined in claim 1 wherein the audio processor means comprises:
means for inputting the audio speech signal and for converting it to an audio electrical signal; and
means for conditioning the audio electrical signal so that it is in a representative electrical form that is suitable for digital sampling.
3. A sound recognition system as defined in claim 2 wherein the conditioning means comprises:
signal amplification means for amplifying the audio electrical signal to a predetermined level;
means for limiting the level of the amplified audio electrical signal to a predetermined output level; and
filter means, connected to the limiting means, for limiting the audio electrical signal to a predetermined maximum frequency of interest and thereby providing the representative audio electrical signal.
4. A sound recognition system as defined in claim 1, further comprising electronic means for receiving at least one word in a preselected language corresponding to the at least one phoneme sound type contained within the audio speech signal, and for programmably processing the at least one word as either a data input or as a command input.
5. A sound recognition system as defined in claim 1, wherein the time domain characteristic includes at least one of the following: an average amplitude of the audio speech signal; an absolute difference average of the audio speech signal; and a zero crossing rate of the audio speech signal.
6. A sound recognition system as defined in claim 1, wherein the at least one frequency domain characteristic includes at least one of the following: a frequency of at least one of said filtered segmentized portions; and an amplitude of at least one of said filtered segmentized portions.
7. A sound recognition system for identifying the phoneme sound types that are contained within an audio speech signal, the sound recognition system comprising:
audio processor means for receiving an audio speech signal and for converting the audio speech signal into a representative audio electrical signal;
analog-to-digital converter means for digitizing the audio electrical signal at a predetermined sampling rate so as to produce a digitized audio signal;
filter means for providing a plurality of filter bands having predetermined high and low cutoff frequencies through which segmentized portions of the digitized audio signal are passed; and
sound recognition means for programmably carrying out the following program steps:
(a) performing a time domain analysis on the segmentized portions of the digitized audio signal so as to identify at least one time domain sound characteristic of said audio speech signal;
(b) filtering the segmentized portions of the digitized audio signal through each of the plurality of filter bands;
(c) measuring at least one frequency domain sound characteristic of each of said filtered segmentized portions; and
(d) based on the at least one time domain characteristic and the at least one frequency domain characteristic, identifying at least one phoneme sound type contained within the audio speech signal.
8. A sound recognition system as defined in claim 7 wherein the audio processor means comprises:
means for inputting the audio speech signal and for converting it to an audio electrical signal; and
means for conditioning the audio electrical signal so that it is in a representative electrical form that is suitable for digital sampling.
9. A sound recognition system as defined in claim 8 wherein the conditioning means comprises:
signal amplification means for amplifying the audio electrical signal to a predetermined level;
means for limiting the level of the amplified audio electrical signal to a predetermined output level; and
filter means, connected to the limiting means, for limiting the audio electrical signal to a predetermined maximum frequency of interest and thereby providing the representative audio electrical signal.
10. A sound recognition system as defined in claim 9, wherein the at least one time domain characteristic includes at least one of the following: an average amplitude of the audio speech signal; an absolute difference average of the audio speech signal; and a zero crossing rate of the audio speech signal.
11. A sound recognition system as defined in claim 10, wherein the at least one frequency domain characteristic includes at least one of the following: a frequency of at least one of said filtered segmentized portions; and an amplitude of at least one of said filtered segmentized portions.
12. A sound recognition system as defined in claim 11, wherein the at least one phoneme sound type contained within the audio speech signal is identified by comparing the at least one measured frequency domain characteristic to a plurality of sound standards each having an associated phoneme sound type and at least one corresponding standard frequency domain characteristic, wherein the at least one identified sound type is the sound standard type having a standard frequency domain characteristic that matches the measured frequency domain characteristic most closely.
13. A sound recognition system as defined in claim 12, wherein the at least one measured frequency domain characteristic, and the plurality of standard frequency domain characteristics are expressed in terms of a chromatic scale.
14. A sound recognition system as defined in claim 13, further comprising electronic means for receiving at least one word in a preselected language corresponding to the at least one phoneme sound type contained within the audio speech signal, and for programmably processing the at least one word as either a data input or as a command input.
15. A sound recognition system for identifying the phoneme sound types that are contained within an audio speech signal, the sound recognition system comprising:
audio processor means for receiving an audio speech signal and for converting the audio speech signal into a representative audio electrical signal;
analog-to-digital converter means for digitizing the audio electrical signal at a predetermined sampling rate so as to produce a digitized audio signal;
filter means for providing a plurality of filter bands having predetermined high and low cutoff frequencies through which segmentized portions of the digitized audio signal are passed;
digital sound processor means for (a) performing a time domain analysis on the segmentized portions of the digitized audio signal so as to identify at least one time domain sound characteristic of said audio speech signal, and for (b) measuring at least one frequency domain sound characteristic of each of the filtered segmentized portions; and
host sound processor means for identifying at least one phoneme sound type contained within the audio speech signal based on the at least one time domain characteristic and the at least one frequency domain characteristic, and for translating said at least one phoneme sound type into at least one representative word of a preselected language.
16. A sound recognition system as defined in claim 15 wherein the audio processor means comprises:
means for inputting the audio speech signal and for converting it to an audio electrical signal; and
means for conditioning the audio electrical signal so that it is in a representative electrical form that is suitable for digital sampling.
17. A sound recognition system as defined in claim 16 wherein the conditioning means comprises:
signal amplification means for amplifying the audio electrical signal to a predetermined level;
means for limiting the level of the amplified audio electrical signal to a predetermined output level; and
filter means, connected to the limiting means, for limiting the audio electrical signal to a predetermined maximum frequency of interest and thereby providing the representative audio electrical signal.
18. A sound recognition system as defined in claim 15, further comprising electronic means for receiving at least one word in a preselected language corresponding to the at least one phoneme sound type contained within the audio speech signal, and for programmably processing the at least one word as either a data input or as a command input.
19. A sound recognition system as defined in claim 15, wherein the said at least one time domain characteristic includes at least one of the following: a average amplitude of the audio speech signal; a absolute difference average of the audio speech signal; and a zero crossing rate of the audio speech signal.
20. A sound recognition system as defined in claim 15, wherein the at least one frequency domain characteristic includes at least one of the following: a frequency of at least one of said filtered segmentized portions; and an amplitude of at least one of said filtered segmentized portions.
21. A sound recognition system as defined in claim 15, wherein the digital sound processor means comprises:
first programmable means for programmably executing a predetermined series of program steps;
program memory means for storing the predetermined series of program steps utilized by said first programmable means; and
data memory means for providing a digital storage area for use by said first programmable means.
22. A sound recognition system as defined in claim 15, wherein the host sound processor means comprises:
second programmable means for programmably executing a predetermined series of program steps;
program memory means for storing the predetermined series of program steps utilized by said second programmable means; and
data memory means for providing a digital storage area for use by said first programmable means.
23. A sound recognition system for identifying the phoneme sound types that are contained within an audio speech signal, the sound recognition system comprising:
audio processor means for receiving an audio speech signal and for converting the audio speech signal into a representative audio electrical signal;
analog-to-digital converter means for digitizing the audio electrical signal at a predetermined sampling rate so as to produce a digitized audio signal;
filter means for providing a plurality of filter bands having predetermined high and low cutoff frequencies through which segmentized portions of the digitized audio signal are passed; and
digital sound processor means for programmably carrying out the following program steps:
(a) performing a time domain analysis on the segmentized portions of the digitized audio signal so as to identify at least one time domain sound characteristic of said audio speech signal;
(b) successively filtering the segmentized portions of the digitized audio signal;
(c) measuring at least one frequency domain sound characteristic from each of said filtered portions; and
host sound processor means for programmably carrying out the following program steps:
(a) based on the at least one time domain characteristic and the at least one frequency domain characteristic, identifying at least one phoneme sound type contained within the audio speech signal; and
(b) translating said at least one phoneme sound type into at least one representative word of a preselected language.
24. A sound recognition system as defined in claim 23 wherein the audio processor means comprises:
means for inputting the audio speech signal and for converting it to an audio electrical signal; and
means for conditioning the audio electrical signal so that it is in a representative electrical form that is suitable for digital sampling.
25. A sound recognition system as defined in claim 24 wherein the conditioning means comprises:
signal amplification means for amplifying the audio electrical signal to a predetermined level;
means for limiting the level of the amplified audio electrical signal to a predetermined output level; and
filter means, connected to the limiting means, for limiting the audio electrical signal to a predetermined maximum frequency of interest and thereby providing the representative audio electrical signal.
26. A sound recognition system as deemed in claim 25, wherein the at least one time domain characteristic includes at least one of the following: an average amplitude of the audio speech signal; an absolute difference average of the audio speech signal; and a zero crossing rate of the audio speech signal.
27. A sound recognition system as defined in claim 26, wherein the said at least one frequency domain characteristic includes at least one of the following: a frequency of at least one of said filtered portions; and an amplitude of at least one of said filtered portions.
28. A sound recognition system as defined in claim 27, wherein the at least one phoneme sound type contained within the audio speech signal is identified by comparing the at least one measured frequency domain characteristic to a plurality of sound standards each having an associated phoneme sound type and at least one corresponding standard frequency domain characteristic, wherein the at least one identified sound type is the sound standard type having a standard frequency domain characteristic that matches the measured frequency domain characteristic most closely.
29. A sound recognition system as defined in claim 28, wherein the at least one measured frequency domain characteristic, and the plurality of standard frequency domain characteristics are expressed in terms of a chromatic scale.
30. A sound recognition system as defined in claim 29, further comprising electronic means for receiving the at least one representative word, and for programmably processing the at least one word as either a data input or as a command input.
31. A method for identifying the phoneme sound types that are contained within an audio speech signal, the method comprising the steps of:
(a) receiving an audio speech signal;
(b) converting the audio speech signal into a representative audio electrical signal;
(c) digitizing the audio electrical signal at a predetermined sampling rate so as to produce a digitized audio signal that is segmentized to form a plurality of separate time sliced signals;
(d) performing a time domain analysis on the digitized audio signal so as to identify at least one time domain sound characteristic of said audio speech signal;
(e) using a plurality of filter bands having predetermined cutoff frequencies to successively filter the time sliced signals of the digitized audio signal;
(f) measuring at least one frequency domain sound characteristic from each of said filtered time sliced signals; and
(g) based on the at least one time domain characteristic and the at least one frequency domain characteristic, identifying at least one phoneme sound type contained within the audio speech signal.
32. A sound recognition system as defined in claim 31, wherein the said at least one time domain characteristic includes at least one of the following: an average amplitude of the audio speech signal; an absolute difference average of the audio speech signal; and a zero crossing rate of the audio speech signal.
33. A sound recognition system as defined in claim 31, wherein the said at least one frequency domain characteristic includes at least one of the following: a frequency of at least one of said filtered time sliced signals; and an amplitude of at least one of said filtered time sliced signals.
34. A sound recognition system as defined in claim 31, wherein the at least one phoneme sound type contained within the audio speech signal is identified by comparing the at least one measured frequency domain characteristic to a plurality of sound standards each having an associated phoneme sound type and at least one corresponding standard frequency domain characteristic, wherein the at least one identified sound type is the sound standard type having a standard frequency domain characteristic that matches the measured frequency domain characteristic most closely.
35. A sound recognition system as defined in claim 34, wherein the at least one measured frequency domain characteristic, and the plurality of standard frequency domain characteristics are expressed in terms of a chromatic scale.
36. A computer program product for use in a computerized sound recognition system that is adapted for receiving an audio speech signal and converting the audio speech signal into a representative audio electrical signal that is digitized, the computer program product comprising:
a computer readable medium for storing computer readable code means which, when executed by the computerized sound recognition system, will enable the system to identify phoneme sound types that are contained within the audio speech signal; and
wherein the computer readable code means is comprised of computer readable instructions for causing the computerized sound recognition system to execute a method comprising the steps of:
performing a time domain analysis on the digitized audio signal so as to identify a plurality of time sound characteristics of said audio speech signal;
performing a frequency domain analysis on the digitized audio signal so as to identify a plurality of frequency domain sound characteristics of said audio speech signal; and
based on the time domain characteristics and the frequency domain characteristics, identifying the phoneme sound types contained within the audio speech signal.
Description
BACKGROUND OF THE INVENTION

1. Technical Field

The present invention relates generally to speech recognition. More particularly, the present invention is directed to a system and method for accurately recognizing continuous human speech from any speaker.

2. Background Information

Linguists, scientists and engineers have endeavored for many years to construct machines that can recognize human speech. Although in recent years this goal has begun to be realized in certain respects, currently available systems have not been able to produce results that even closely emulate human performance. This inability to provide satisfactory speech recognition is due primarily to the difficulties that are involved in extracting and identifying the individual sounds that make up human speech. These difficulties are exacerbated by the fact there are such wide acoustic variations that occur between different speakers.

Simplistically, speech may be considered as a sequence of sounds taken from a set of forty or so basic sounds called "phonemes." Different sounds, or phonemes, are produced by varying the shape of the vocal tract through muscular control of the speech articulators (lips, tongue, jaw, etc.). A stream of a particular set of phonemes will collectively represent a word or a phrase. Thus, extraction of the particular phonemes contained within a speech signal is necessary to achieve voice recognition.

However, a number of factors are present that make phoneme extraction extremely difficult. For instance, wide acoustic variations occur when the same phoneme is spoken by different speakers. This is due to the differences in the vocal apparatus, such as the vocal-tract length. Moreover, the same speaker may produce acoustically different versions of the same phoneme from one rendition to the next. Also, there are often no identifiable boundaries between sounds or even words. Other difficulties result from the fact that phonemes are spoken with wide variations in dialect, intonation, rhythm, stress, volume, and pitch. Finally, the speech signal may contain wide variations in speech-related noises that make it difficult to accurately identify and extract the phonemes.

The speech recognition devices that are currently available attempt to minimize the above problems and variations by providing only a limited number of functions and capabilities. For instance, many existing systems are classified as "speaker-dependent" systems. A speaker-dependent system must be "trained" to a single speaker's voice by obtaining and storing a database of patterns for each vocabulary word uttered by that particular speaker. The primary disadvantage of these types of systems is that they are "single speaker" systems, and can only be utilized by the speaker who has completed the time consuming training process. Further, the vocabulary size of such systems is limited to the specific vocabulary contained in the database. Finally, these systems typically cannot recognize naturally spoken continuous speech, and require the user to pronounce words separated by distinct periods of silence.

Currently available "speaker-independent" systems are also severely limited in function. Although any speaker can use the system without the need for training, these systems can only recognize words from an extremely small vocabulary. Further, they too require that the words be spoken in isolation with distinct pauses between words, and thus cannot recognize naturally spoken continuous speech.

OBJECTS AND BRIEF SUMMARY OF THE INVENTION

The present invention has been developed in response to the present state of the art, and in particular, in response to these and other problems and needs that have not been fully or completely solved by currently available solutions for speech recognition. It is therefore a primary object of the present invention to provide a novel system and method for achieving speech recognition.

Another object of the present invention is to provide a speech recognition system and method that is user independent, and that can thus be used to recognize speech utterances from any speaker of a given language.

A related object of the present invention is to provide a speech recognition system and method that does not require a user to first "train" the system with the user's individual speech patterns.

Yet another object of the present invention is to provide a speech recognition system and method that is capable of receiving and processing an incoming speech signal in substantially real time, thereby allowing the user to speak at normal conversational speeds.

A related object of the present invention is to provide a speech recognition system and method that is capable of accurately extracting various sound characteristics from a speech signal, and then converting those sound characteristics into representative phonemes.

Still another object of the present invention is to provide a speech recognition system and method that is capable of converting a stream of phonemes into an intelligible format.

Another object of the present invention is to provide a speech recognition system and method that is capable of performing speech recognition on a substantially unlimited vocabulary.

These and other objects and features of the present invention will become more fully apparent from the following description and appended claims, or may be learned by the practice of the invention as set forth hereinafter.

Briefly summarized, the foregoing and other objects are achieved with a novel speech recognition system and method, which can accurately recognize, continuous speech utterances from any speaker of a given language. In the preferred embodiment, an audio speech signal is received from a speaker and input to an audio processor means. The audio processor means receives the speech signal, converts it into a corresponding electrical format, and then electrically conditions the signal so that it is in a form that is suitable for subsequent digital sampling.

Once the audio speech signal has been converted to a representative audio electrical signal, it is sent to an analog-to-digital converter means. The A/D converter means samples the audio electrical signal at a suitable sampling rate, and outputs a digitized audio signal.

The digitized audio signal is then programmably processed by a sound recognition means, which processes the digitized audio signal in a manner so as to extract various time domain and frequency domain sound characteristics, and then identify the particular phoneme sound type that is contained within the audio speech signal. This characteristic extraction and phoneme identification is done in a manner such that the speech recognition occurs regardless of the source of the audio speech signal. Importantly, there is no need for a user to first "train" the system with his or her individual voice characteristics. Further, the process occurs in substantially real time so that the speaker is not required to pause between each word, and can thus speak at normal conversational speeds.

In addition to extracting phoneme sound types from the incoming audio speech signal, the sound recognition means implements various linguistic processing techniques to translate the phoneme string into a corresponding word or phrase. This can be done for essentially any language that is made up of phoneme sound types.

In the preferred embodiment, the sound recognition means is comprised of a digital sound processor means and a host sound processor means. The digital sound processor includes a programmable device and associated logic to programmably carry out the program steps used to digitally process the audio speech signal, and thereby extract the various time domain and frequency domain sound characteristics of that signal. This sound characteristic data is then stored in a data structure, which corresponds to the specific portion of the audio signal.

The host sound processor means also includes a programmable device and its associated logic. It is programmed to carry out the steps necessary to evaluate the various sound characteristics contained within the data structure, and then generate the phoneme sound type that corresponds to those particular characteristics. In addition to identifying phonemes, in the preferred embodiment the host sound processor also performs the program steps needed to implement the linguistic processing portion of the overall method. In this way, the incoming stream of phonemes are translated to the representative word or phrase.

The preferred embodiment further includes an electronic means, connected to the sound recognition means, for receiving the word or phrase translated from the incoming stream of identified phonemes. The electronic means, as for instance a personal computer, then programmably processes the word as either data input, as for instance text to a wordprocessing application, or as a command input, as for instance an operating system command.

BRIEF DESCRIPTION OF THE DRAWINGS

In order that the manner in which the above-recited and other advantages and objects of the invention are obtained, a more particular description of the invention briefly described above will be rendered by reference to a specific embodiment thereof which is illustrated in the appended drawings. Understanding that these drawings depict only a typical embodiment of the invention and are not to be considered to be limiting of its scope, the invention in its presently understood best mode will be described and explained with additional specificity and detail through the use of the accompanying drawings in which:

FIG. 1 is functional block diagram of the overall speech recognition system;

FIG. 2 is a more detailed functional block diagram illustrating the speech recognition system;

FIGS. 3A-H, J-N, P-Y are a schematic illustrating in detail the circuitry that makes up the functional blocks in FIG. 2;

FIG. 4 is a functional flow-chart illustrating the overall program method of the present invention;

FIGS. 5A-5B is a flow-chart illustrating the program method used to implement one of the functional blocks of FIG. 4;

FIGS. 6-6D is a flow-chart illustrating the program method used to implement one of the functional blocks of FIG. 4;

FIG. 7 is a flow-chart illustrating the program method used to implement one of the functional blocks of FIG. 4;

FIGS. 8-8C is a flow-chart illustrating the program method used to implement one of the functional blocks of FIG. 4;

FIG. 9 is a flow-chart illustrating the program method used to implement one of the functional blocks of FIG. 4;

FIGS. 10-10C is a flow chart illustrating the program method used to implement one of the functional blocks of FIG. 4;

FIGS. 11, 12 are flow-charts illustrating the program method used to implement one of the functional blocks of FIG. 4;

FIGS. 12A-12C are x-y plots of example standard sound data.

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

The following detailed description is divided into two parts. In the first part the overall system is described, including a detailed description of the functional blocks which make up the system, and the manner in which the various functional blocks are interconnected. In part two, the method by which the overall system is programmably controlled to achieve real-time, user-independent speech recognition is described.

I. THE SYSTEM

Reference is first made to FIG. 1, where one presently preferred embodiment of the overall speech recognition system is designated generally at 10. The system 10 includes an audio processor means for receiving an audio speech signal and for converting that signal into a representative audio electrical signal. In the preferred embodiment, the audio processor means is comprised of a means for inputting an audio signal and converting it to an electrical signal, such as a standard condenser microphone shown generally at 12. Various other input devices could also be utilized to input an audio signal, including, but not limited to such devices as a dictaphone, telephone or a wireless microphone.

In addition to microphone 12, the audio processor means also preferably comprises additional appropriate audio processor circuitry 14. This circuitry 14 receives the audio electrical signal generated by the microphone 12, and then functions so as to condition the signal so that it is in a suitable electrical condition for digital sampling.

The audio processor circuitry 14 is then electrically connected to analog-to-digital converter means, illustrated in the preferred embodiment as A/D conversion circuitry 34. This circuitry 34 receives the audio electrical signal, which is in an analog format, and converts it to a digital format, outputting a digitized audio signal.

This digitized audio signal is then passed to a sound recognition means, which in the preferred embodiment corresponds to the block designated at 16 and referred to as the sound recognition processor circuitry. Generally, the sound recognition processor circuitry 16 programmably analyzes the digitized version of the audio signal in a manner so that it can extract various acoustical characteristics from the signal. Once the necessary characteristics are obtained, the circuitry 16 can identify the specific phoneme sound types contained within the audio speech signal. Importantly, this phoneme identification is done without reference to the speech characteristics of the individual speaker, and is done in a manner such that the phoneme identification occurs in real time, thereby allowing the speaker to speak at a normal rate of conversation.

The sound recognition processor circuitry 16 obtains the necessary acoustical characteristics in two ways. First, it evaluates the time domain representation of the audio signal, and from that representation extracts various parameters representative of the type of phoneme sound contained within the signal. The sound type would include, for example, whether the sound is "voiced," "unvoiced," or "quiet."

Secondly, the sound recognition processor circuitry 16 evaluates the frequency domain representation of the audio signal. Importantly, this is done by successively filtering the time domain representation of the audio signal using a predetermined number of filters having a various cutoff frequencies. This produces a number of separate filtered signals, each of which are representative of an individual signal waveform which is a component of the complex audio signal waveform. The sound recognition processor circuitry 16 then "measures" each of the filtered signals, and thereby extracts various frequency domain data, including the frequency and amplitude of the of the signals. These frequency domain characteristics, together with the time domain characteristics, provide sufficient "information" about the audio signal such that the processor circuitry 16 can identify the phoneme sounds that are contained therein.

Once the sound recognition processor circuitry 16 has extracted the corresponding phoneme sounds, it programmably invokes a series of linguistic program tools. In this way, the processor circuitry 16 translates the series of identified phonemes into the corresponding syllable, word or phrase.

With continued reference to FIG. 1, electrically connected to the sound recognition processor circuitry 16 is a host computer 22. In one preferred embodiment, the host computer 22 is a standard desktop personal computer, however it could be comprised of virtually any device utilizing a programmable computer that requires data input and/or control. For instance, the host computer 22 could be a data entry system for automated baggage handling, parcel sorting, quality control, computer aided design and manufacture, and various command and control systems.

As the processor circuitry 16 translates the phoneme string, the corresponding word or phrase is passed to the host computer 22. The host computer 22, under appropriate program control, then utilizes the word or phrase as an operating system or application command or, alternatively, as data that is input directly into an application, such as a wordprocessor or database.

Reference is next made to FIG. 2 where one presently preferred embodiment of the voice recognition system 10 is shown in further detail. As is shown, an audio speech signal is received at microphone 12, or similar device. The representative audio electrical signal is then passed to the audio processor circuitry 16 portion of the system. In the preferred embodiment of this circuit, the audio electrical signal is input to a signal amplification means for amplifying the signal to a suitable level, such as amplifier circuit 26. Although a number of different circuits could be used to implement this function, in the preferred embodiment, amplifier circuit 26 consists of a two stage operational amplifier configuration, arranged so as to provide an overall gain of approximately 300. With such a configuration, with a microphone 12 input of approximately 60 dbm, the amplifier circuit 26 will produce an output signal at approximately line level.

In the preferred embodiment, the amplified audio electrical signal is then passed to a means for limiting the output level of the audio signal so as to prevent an overload condition to other components contained within the system 10. The limiting means is comprised of a limiting amplifier circuit 28, which can be designed using a variety of techniques, one example of which is shown in the detailed schematic of FIG. 3.

Next, the amplified audio electrical signal is passed to a filter means for filtering high frequencies from the electrical audio signal, as for example anti-aliasing filter circuit 30. This circuit, which again can be designed using any one of a number of circuit designs, merely limits the highest frequency that can be passed on to other circuitry within the system 10. In the preferred embodiment, the filter circuit 30 limits the signal's frequency to less than about 12 kHz.

The audio electrical signal, which is in an analog format, is then passed to a analog-to-digital converter means for digitizing the signal, which is shown as A/D conversion circuit 34. In the preferred embodiment, A/D conversion circuit 34 utilizes a 16-bit analog to digital converter device, which is based on Sigma-Delta sampling technology. Further, the device must be capable of sampling the incoming analog signal at a rate sufficient to avoid aliasing errors. At a minimum, the sampling rate should be at least twice the incoming sound wave's highest frequency (the Nyquest rate), and in the preferred embodiment the sampling rate is 44.1 kHz. It will be appreciated that any one of a number of A/D conversion devices that are commercially available could be used. A presently preferred component, along with the various support circuitry, is shown in the detailed schematic of FIG. 3.

With continued reference to FIG. 2, having converted the audio electrical signal to a digital form, the digitized signal is next supplied to the sound recognition processor circuitry 16. In the presently preferred embodiment, the sound recognition processor circuitry 16 is comprised of a digital sound processor means and a host sound processor means, both of which are preferably comprised of programmable devices. It will be appreciated however that under certain conditions, the sound recognition processor circuitry 16 could be comprised of suitable equivalent circuitry which utilizes a single programmable device.

In the presently preferred embodiment, the digital sound processor means is comprised of the various circuit components within the dotted box 18 and referred to as the digital sound processor circuitry. This circuitry receives the digitized audio signal, and then programmably manipulates that data in a manner so as to extract various sound characteristics. Specifically, the circuitry 18 first analyzes the digitized audio signal in the time domain and, based on that analysis, extracts at least one time domain sound characteristic of the audio signal. The time domain characteristics of interest help determine whether the audio signal contains a phoneme sound that is "voiced," "unvoiced," or "quiet."

The digital sound processor circuitry 18 also manipulates the digitized audio signal so as to obtain various frequency domain information about the audio signal. This is done by filtering the audio signal through a number of filter bands and generating a corresponding number of filtered signals, each of which are still in time domain. The circuitry 18 measures various properties exhibited by these individual waveforms, and from those measurements, extracts at least one frequency domain sound characteristic of the audio signal. The frequency domain characteristics of interest include the frequency, amplitude and slope of each of the component signals obtained as a result of the filtering process. These characteristics are then stored and used to determine the phoneme sound type that is contained in the audio signal.

With continued reference to FIG. 2, the digital sound processor circuitry 18 is shown as preferably comprising a first programmable means for analyzing the digitized audio signal under program control, such as digital sound processor 36. Digital sound processor 36 is preferably a programmable, 24-bit general purpose digital signal processor device, such as the Motorola DSP56001. However, any one of a number of commercially available digital signal processors could also be used.

As is shown, digital sound processor 36 is preferably interfaced--via a standard address, data and control bus-type arrangement 38--to various other components. They include: a program memory means for storing the set of program steps executed by the DSP 36, such as DSP program memory 40; data memory means for storing data utilized by the DSP 36, such as DSP data memory 42; and suitable control logic 44 for implementing the various standard timing and control functions such as address and data gating and mapping. It will be appreciated by one of skill in the art that various other components and functions could be used in conjunction with the digital sound processor 36.

With continued reference to FIG. 2, in the presently preferred embodiment, the host sound processor means is comprised of the various circuit components within the dotted box 20 and referred to as the host sound processor circuitry. This host sound processor circuitry 20 is electrically connected and interfaced, via an appropriate host interface 52, to the digital sound processor circuitry 18. Generally, this circuitry 20 receives the various audio signal characteristic information generated by the digital sound processor circuitry 18 via the host interface 52. The host sound processor circuitry 20 analyzes this information and then identifies the phoneme sound type(s) that are contained within the audio signal by comparing the signal characteristics to standard sound data that has been compiled by testing a representative cross-section of speakers. Having identified the phoneme sounds, the host sound processor circuitry 20 utilizes various linguistic processing techniques to translate the phonemes into a representative syllable, word or phrase.

The host sound processor circuitry 20 is shown as preferably comprising a second programmable means for analyzing the digitized audio signal characteristics under program control, such as host sound processor 54. Host sound processor 36 is preferably a programmable, 32-bit general purpose CPU device, such as the Motorola 68EC030. However, any one of a number of commercially available programmable processors could also be used.

As is shown, host sound processor 54 is preferably interfaced--via a standard address, data and control bus-type arrangement 56--to various other components. They include: a program memory means for storing the set of program steps executed by the host sound processor 54, such as host program memory 58; data memory means for storing data utilized by the host sound processor 54, such as host data memory 60; and suitable control logic 64 for implementing the various standard timing and control functions such as address and data gating and mapping. Again, it will be appreciated by one of skill in the art that various other components and functions could be used in conjunction with the host sound processor 54.

Also included in the preferred embodiment is a means for interfacing the host sound processor circuitry 20 to an external electronic device. In the preferred embodiment, the interface means is comprised of standard RS-232 interface circuitry 66 and associated RS-232 cable 24. However, other electronic interface arrangements could also be used, such as a standard parallel port interface, a musical instrument digital interface (MIDI), or a non-standard electrical interface arrangement.

In the preferred embodiment, the host sound processor circuitry 20 is interfaced to a electronic means for receiving the word generated by the host sound processor circuitry 20 and for processing that word as either a data input or as a command input. By way of example and not limitation, the electronic receiving means is comprised of a host computer 22, such as a standard desktop personal computer. The host computer 22 is connected to the host sound processor circuitry 20 via the RS-232 interface 66 and cable 24 and, via an appropriate program method, utilizes incoming words as either data, such as text to a wordprocessor application, or as a command, such as to an operating system or application program. It will be appreciated that the host computer 22 can be virtually any electronic device requiring data on command input.

One example of an electronic circuit which has been constructed and used to implement the above described block diagram is illustrated in FIGS. 3A-3Y. These figures are a detailed electrical schematic diagram showing the interconnections, part number and/or value of each circuit element used. It should be noted that FIGS. 3A-3Y are included merely to show an example of one such circuit which has been used to implement the functional blocks described in FIG. 2. Other implementations could be designed that would also work satisfactorily.

II. The Method

Referring now to FIG. 4, illustrated is a functional flow chart showing one presently preferred embodiment of the overall program method used by the present system. As is shown, the method allows the voice recognition system 10 to continuously receive an incoming speech signal, electronically process and manipulate that signal so as to generate the phonetic content of the signal, and then produce a word or stream of words that correspond to that phonetic content. Importantly, the method is not restricted to any one speaker, or group of speakers. Rather, it allows for the unrestricted recognition of continuous speech utterances from any speaker of a given language.

Following is a general description of the overall functions carried out by the present method. A more detailed description of the preferred program steps used to carry out these functions will follow. Referring first to the functional block indicated at 100, the audio processor 16 portion of the system receives the audio speech signal at microphone 12, and the A/D conversion circuit 34 digitizes the analog signal at a suitable sampling rate. The preferred sampling rate is 44.1 kHz, although other sampling rates could be used, as long as it complies with the Nyquist sampling rate so as to avoid aliasing problems. This digitized speech signal is then broken-up into successive "time segments." In the preferred embodiment, each of these time segments contains 10,240 data points, or 232 milliseconds of time domain data.

Each time segment of 10,240 data points is then passed to the portion of the algorithm labeled "Evaluate Time Domain," shown at numeral 102. This portion of the method further breaks the time segments up into successive "time slices." Each time slice contains 256 data points, or 5.8 milliseconds of time domain data. Various sound characteristics contained within each time slice are then extracted. Specifically, in the preferred embodiment the absolute average envelope amplitude, the absolute difference average, and the zero crossing rate for the portion of the speech signal contained within each time slice is calculated and stored in a corresponding data structure. From these various characteristics, it is then determined whether the particular sound contained within the time slice is quiet, voiced or unvoiced. This information is also stored in the time slice's corresponding data structure.

The next step in the overall algorithm is shown at 104 and is labeled "Decompose." In this portion of the program method, each time slice is broken down into individual component waveforms by successively filtering the time slice using a plurality of filter bands. From each of these filtered signals, the Decompose function directly extracts additional sound identifying characteristics by "measuring" each signal. Identifying characteristics include, for example, the fundamental frequency of the time slice if voiced; and the frequency and amplitude of each of the filtered signals. This information is also stored in each time slice's corresponding data structure. The next step in the overall algorithm is at 106 and is labeled "Point of Maximum Intelligence." In this portion of the program, those time slices that Contain sound data which is most pertinent to the identification of the sound(s) are identified as points of "maximum intelligence;" the other time slices are ignored. In addition to increasing the accuracy of subsequent phoneme identification, this function also reduces the amount of processing overhead required to identify the sound(s) contained within the time segment.

Having identified those time slices that are needed to identify the particular sound(s) contained within the time segment, the system then executes the program steps corresponding to the functional block 110 labeled "Evaluate." In this portion of the algorithm, all of the information contained within each time slice's corresponding data structure is analyzed, and up to five of the most probable phonetic sounds (i.e., phonemes) contained within the time slice are identified. Each possible sound is also assigned a probability level, and are ranked in that order. The identified sounds and their probabilities are then stored within the particular time slice's data structure. Each individual phoneme sound type is identified by way of a unique identifying number referred to as a "PASCII" value.

The next functional step in the overall program method is performed by the system at the functional block 110 labeled "Compress Phones." In this function, the time slices that do not correspond to "points of maximum intelligence" are discarded. Only those time slices which contain the data necessary to identify the particular sound are retained. Also, time slices which contain contiguous "quiet" sections are combined, thereby further reducing the overall number of time slices. Again, this step reduces the amount of processing that must occur and further facilitates real time sound recognition.

At this point in the algorithm, there remains a sequence of time slices, each of which has a corresponding data structure containing various sound characteristics culled from both the time domain and the frequency domain. Each structure also identifies the most probable phoneme sound type corresponding to those particular sound characteristics. This data is passed to the next step of the overall program method, shown at functional block 112 and labeled "Linguistic Processor." The Linguistic processor receives the data structures, and translates the sound stream (i.e., stream of phonemes) into the corresponding English letter, syllable, word or phrase. This translation is generally accomplished by performing a variety of linguistic processing functions that match the phonemic sequences against entries in the system lexicon. The presently preferred linguistic functions include a phonetic dictionary look-up, a context checking function and database, and a basic grammar checking function.

Once the particular word or phrase is identified, it is passed to the "Command Processor" portion of the algorithm, as shown at functional block 114. The Command processor determines whether the word or phrase constitutes text that should be passed as data to a higher level application, such as a wordprocessor, or whether it constitutes a command that is to be passed directly to the operating system or application command interface.

As has been noted in the above general description, a data structure is preferably maintained for each time slice of data (i.e., 256 samples of digitized sound data; 5.8 milliseconds of sound) within system memory. This data structure is referred to herein as the "Higgins" structure, and its purpose is to dynamically store the various sound characteristics and data that can be used to identify the particular phoneme type contained within the corresponding time slice. Although other information could also be stored in the Higgins structure, TABLE I illustrates one preferred embodiment of the its contents. The data structure and its contents will be discussed in further detail below.

              TABLE 1______________________________________VARIABLE NAME        CONTENTS______________________________________TYPE         Whether sound is voiced, unvoiced,        quiet or Not processed.LOCATION     Array location of where Time Slice starts.SIZE         Number of sample data points in Time        Slice.Ls      Average amplitude of signal in time        domain.fo      Fundamental Frequency of signal.FFREQ        Array containing the frequency of each        filtered signal contained in time slice.AMPL         Array containing the amplitude of each        filtered signal.ZCR     Zero Crossing Rate of signal        in time domain.PMI          Variable indicating maximum formant        stability; value indicates duration.sumSlope     Sum of absolute values of filtered        signal slopes.POSSIBLE     Array containing up to five mostPHONEMES     probable phonemes        contained in time slice, including        for each phoneme: confidence level,        standard for relative amplitude, standard        for ZCR, Standard for duration        for phoneme.______________________________________

The various steps used to accomplish the method illustrated in FIG. 4 will now be discussed in more detail by making specific reference to one presently preferred embodiment of the invention. It should be appreciated that the particular program steps which are illustrated in the detailed flow charts contained in FIGS. 5 through 11 are intended merely as an example of the presently preferred embodiment and the presently understood best mode of implementing the overall functions which are represented by the flow chart of FIG. 4.

Referring first to FIG. 5A, the particular program steps corresponding to the "Evaluate Time Domain" function illustrated in functional block 102 of FIG. 4 are shown. As already noted, the Audio Processor 16 receives an audio speech signal from the microphone 12. The A/D conversion circuitry 34 then digitally samples that signal at a predetermined sampling rate, such as the 44.1 kHz rate used in the preferred embodiment. This time domain data is divided into separate, consecutive time segments of predetermined lengths. In the preferred embodiment, each time segment is 232 milliseconds in duration, and consists of 10,240 digitized data points. Each time segment is then passed, one at a time, to the Evaluate Time Domain function, as is shown at step 116 in FIG. 5A. Once received, the time segment is further segmented into a predetermined number of equal "slices" of time. In the preferred embodiment, there are forty of these "time slices" for each time segment, each of which are comprised of 256 data points, or 5.8 milliseconds of speech.

The digital sound processor 36 then enters a program loop, beginning with step 118. As is indicated at that step, for each time slice the processor 36 extracts various time-varying acoustic characteristics. For example, in the preferred embodiment the DSP 36 calculates the absolute average of the amplitude of the time slice signal (LS), the absolute difference average (LD) of the time slice signal and the zero crossing rate (ZCR) Of the time slice signal. The absolute average of the amplitude LS corresponds to the absolute value of the average of the amplitudes (represented as a line level signal voltage) of the data points contained within the time slice. The absolute difference average LD is the average amplitude difference between the data points in the time slice (i.e., calculated by taking the average of the differences between the absolute value of one data point's amplitude to the next data point's). The zero crossing rate ZCR is calculated by dividing the number of zero crossings that occur within the time slice by the number of data points (256) and multiplying the result by 100. The number of zero crossings is equal to the number of times the time domain data crosses the X-axis, whether that crossing be positive-to-negative or negative-to-positive.

The magnitudes of these various acoustical properties can be used to identify the general type of sound contained within each time slice. For instance, the energy of "voiced" speech sounds is generally found at lower frequencies than for "unvoiced" sounds, and the amplitude of unvoiced sounds is generally much lower than the amplitude of voiced sounds. These generalizations are true of all speakers, and general ranges have been identified by analyzing speech data taken from a wide variety of speakers (i.e., men, women, and children). By comparing the various acoustical properties to these predetermined ranges, the sound type can be determined, independent of the particular speaker.

Thus, based on the acoustical properties identified in the previous step, the DSP 36 next proceeds to that portion of the program loop that identifies what type of sound is contained within the particular time slice. In the preferred embodiment, this portion of the code determines, based on previously identified ranges obtained from test data, whether the sound contained within the time slice is "quiet," "voiced" or "unvoiced."

At step 120, the absolute average of the amplitude LS is first compared with a predetermined "quiet level" range, or "QLEVEL" (i.e., an amplitude magnitude level that corresponds to silence). In the preferred embodiment, QLEVEL is equal to 250, but the value can generally be anywhere between 200 and 500. It will be appreciated that the particular "quiet level" may vary depending on the application or environment (e.g., high level of background noise, high d.c. offset present in the A/D conversion or where the incoming signal is amplified to a different level), and thus may be a different value. If LS is less than QLEVEL, the sound contained within the time slice is deemed to be "quiet," and the processor 36 proceeds to step 122. At step 122, the DSP 36 begins to build the Higgins data structure for the current time slice within DSP data memory 42. Here, the processor 36 places an identifier "Q" into a "type" flag of the Higgins data structure for this time slice.

If however, Ls is greater than QLEVEL, then the sound contained within the time slice is not quiet, and the processor 36 proceeds to step 124 to determine whether the sound is instead a "voiced" sound. To make this determination, the zero crossing rate ZCR is first compared with a predetermined crossing-rate value found to be indicative of a voiced sound for most speakers. A low zero-crossing rate implies a low frequency and, in the preferred embodiment, if it is less than or equal to about 10, the speech sound is probably voiced.

If the ZCR does fall below 10, another acoustical property of the sound is evaluated before the determination is made that the sound is voiced. This property is checked by calculating the ratio of LD to LS, and then comparing that ratio to another predetermined value that corresponds to a cut-off point corresponding to voiced sounds in most speakers. In the preferred embodiment, if LD /LS is less than or equal to about 15, then the signal is probably voiced. Thus, if at step 124 it is determined that ZCR is less than or equal to 10 and that LD /LS is less than or equal to about 15, then the sound is deemed to be a voiced type of sound (e.g., the sounds /U/, /d/, /w/, /i/, /e/, etc.). If voiced, the processor 36 proceeds to step 126 and places an identifier "V" into the "type" flag of the Higgins data structure corresponding to that time slice.

If not voiced, then the processor 36 proceeds to program step 120 to determine if the sound is instead "unvoiced," again by comparing the properties identified at step 118 to ranges obtained from user-independent test data. To do so, processor 36 determines whether ZCR is greater than or equal to about 20 and whether LD /LS is greater than or equal to about 30. If both conditions exist, the sound is considered to be an unvoiced type of sound (e.g., certain aspirated sounds). If unvoiced, the processor 36 proceeds to step 130 and places an identifier "U" into the "type" flag of the Higgins data structure for that particular time slice.

Some sounds will fall somewhere between the conditions checked for in steps 124 and 128 (i.e., ZCR falls somewhere between about 11 and 19, and LD /LS falls somewhere between about 16 and 29) and other sound properties must be evaluated to determine whether the sound is voiced or unvoiced. This portion of the program method is performed, as is indicated at step 132, by executing another set of program steps referred to as "Is it Voiced." The programs steps corresponding to this function are illustrated in FIG. 5B, to which reference is now made.

After receiving the current time slice data at step 141, the processor proceeds to step 142, where a digital low pass filter is programmably implemented within the DSP 36. The speech signal contained within the current time slice is then passed through this filter. In the preferred embodiment, the filter removes frequencies above 3000 Hz, and the zero crossing rate, as discussed above, is recalculated. This is because certain voiced fricatives have high frequency noise components that tend to raise the zero crossing rate of the signal. For these types of sounds, elimination of the high frequency components will drop the ZCR to a level which corresponds to other voiced sounds. In contrast, if the sound is an unvoiced fricative, then the ZCR will remain largely unchanged and stay at a relatively high level, because the majority of the signal resides at higher frequencies.

Once the new ZCR has been calculated, program step 144 is performed to further evaluate whether the sound is a voiced or an unvoiced fricative. Here, the time slice's absolute minimum amplitude point is located. Once located, the processor 36 computes the slope (i.e., the first derivative) of the line defined between that point and another data point on the waveform that is located a predetermined distance from the minimum point. In the preferred embodiment, that predetermined distance is 50 data points, but other distance values could also be used. For a voiced fricative sound, the slope will be relatively high since the signal is periodic, and thus exhibits a fairly significant change in amplitude. In contrast, for an unvoiced fricative sound the slope will be relatively low because the signal is not periodic and, having been filtered, will be comprised primarily of random noise having a fairly constant amplitude.

Having calculated the ZCR and the slope, the processor 36 proceeds to step 146 and compares the magnitudes to predetermined values corresponding to the threshold of a voiced fricative for most speakers. In the preferred embodiment, if ZCR is less than about 8, and if the slope is greater than about 35, then the sound contained within the time slice is deemed to be voiced, and the corresponding "true" flag is set at step 150. Otherwise, the sound is considered unvoiced, and the "false" flag is set at step 148. Once the appropriate flag is set, the "Is it Voiced" program sequence returns to its calling routine at step 132, shown in FIG. 5A.

Referring again to FIG. 5A at step 134, based on the results of the previous step 132, the appropriate identifier "U" or "V" is placed into the "type" flag of the data structure for that particular time slice. Once it has been determined whether the speech sound contained within the particular time slice is voiced, unvoiced or quiet, and the Higgins data structure has been updated accordingly at steps 122, 126, 130 or 134, the DSP 36 proceeds to step 136 and determines whether the last of the 256 time slices for this particular time segment has been processed. If so, the DSP 36 returns to the main calling routine (illustrated in FIG. 4) as is indicated at step 140. Alternatively, the DSP 36 obtains the next time slice at step 138, and proceeds as described above.

Referring again to FIG. 4, once the "Evaluate Time Domain Parameters" function shown at functional block 102 has been completed, the "Decompose a Speech Signal" portion of the algorithm shown at functional block 104 is performed.

As will be appreciated from the following description, to accurately identify the sound(s) contained within the time segment, additional identifying characteristics must be culled from the signal. Such characteristics relate to the amplitude and frequency each of the various component signals that make up the complex waveform contained within the time slice. This information is obtained by successively filtering the time slice into its various component signals. Previously, this type of "decomposition" was usually accomplished by performing a Fast Fourier Transform on the sound signal. However, this standard approach is not adequate for evaluating user-independent speech in real time. For many sounds, accurate identification of the individual component frequencies is very difficult, if not impossible, due to the spectral leakage that is inherently present in the FFT's output. Also, because the formant signals contained in speech signals are amplitude modulated due to the glottal spectrum dampening and because most speech signals are non-periodic then, by definition, the FFT is an inadequate tool. However, such information is critical to accomplish user-independent speech recognition with the required level in confidence.

To avoid this problem, in the preferred embodiment of the Decompose a Speech Signal algorithm, a FFT is not performed. Instead, the DSP 36 filters the time slice signal into various component filtered signals. As will be described in further detail, frequency domain data can be extracted directly from each of these filtered signals. This data can then be used to determine the characteristics of the specific phoneme contained within the time slice.

By way of example and not limitation, the detailed program steps used to perform this particular function are shown in the flow chart illustrated in FIG. 6. Referring first to program step 152, the current time segment (10,240 data samples; 232 milliseconds in duration) is received. The program then enters a loop, beginning with step 154, wherein the speech signal contained within the current time segment is successively filtered into its individual component waveforms by using a set of digital bandpass filters having specific frequency bands. In the preferred embodiment, these frequency bands are precalculated, and stored in DSP program memory 40. At step 154, the processor 36 obtains the first filter band, designated as a low frequency (fL) and a high frequency (fH), from this table of predetermined filter cutoff frequencies. In the preferred embodiment, the filter cutoff frequencies are located at: 0 Hz, 250 Hz, 500 Hz, 1000 Hz, 1500 Hz, 2000 Hz, 2500 Hz, 3000 Hz, 3500 Hz, 4000 Hz, 4500 Hz, 5000 Hz, 6000 Hz, 7000 Hz, 8000 Hz, 9000 Hz, and 10,000 Hz. It will be appreciated that different or additional cutoff frequencies could also be used.

Thus, during the first pass through the loop beginning at step 154, fL will be set to 0 Hz, and fH to 250 Hz. The second pass through the loop will set fL to 250 Hz and fH to 500 Hz, and so on.

Having set the appropriate digital filter parameters, the processor 36 then proceeds to step 158, where the actual filtering of the time segment occurs. To do so, this step invokes another function referred to as "Do Filter Pass," which is shown in further detail in FIG. 6A and to which reference is now made.

At step 168 of function Do Filter Pass, the previously calculated filter parameters, as well as the time segment data is received (10,240 data points). At step 170, the coefficients for the filter are obtained from a predetermined table of coefficients that correspond to each of the different filter bands. Alternatively, the coefficients could be recalculated by the processor 36 for each new filter band.

Having set the filter coefficients, the processor 36 executes program step 172, where the current time segment is loaded into the digital filter. Optionally, rather than loading all data samples, the signal may be decimated and only every nth point loaded, where n is in the range of one to four. Before the signal is decimated, it should be low pass filtered down to a frequency less than or equal to the original sample rate divided by 2*n. At step 174. the filtering operation is performed on the current time segment data. The results of the filtering operation are written into corresponding time segment data locations within DSP data memory 42. Although any one of a variety of different digital filter implementations could be used to filter the data, in the preferred embodiment the digital bandpass filter is an IIR cascade-type filter with a Butterworth response.

Once the filtering operation is complete for the current filter band, the processor 36 proceeds to step 176 where the results of the filtering operation are evaluated. This is performed by the function referred to as "Evaluate Filtered Data," which is shown in further detail in FIG. 6B, to which reference is now made.

At step 182 of Evaluate Filtered Data, a time slice of the previously filtered time segment is received. Proceeding next to step 183, the amplitude of this filtered signal is calculated. The amplitude is calculated using the following equation: ##EQU1## where max=the highest amplitude value in the time slice; and min=the lowest amplitude value in the time slice.

At step 184 the frequency of the filtered signal is measured. This is performed by a function called "Measure Frequency of a Filtered Signal," which is shown in further detail in FIG. 6C. Referring to that figure, at step 192 the filtered time slice data is received. At step 194, the processor 36 calculates the slope (i.e., the first derivative) of the filtered signal at each data point. This slope is calculated with reference to the line formed by the previous data point, the data point for which the slope is being calculated, and the data point following it, although other methods could also be used.

Proceeding next to step 196, each of the data point locations corresponding to a slope changing from a positive value to a negative value is located. Zero crossings are determined beginning at the maximum amplitude value in the filtered signal and proceeding for at least three zero crossings. The maximum amplitude value represents the closure of the vocal folds. Taking this frequency measurement after the close of the vocal folds insures the most accurate frequency measurement. At step 198 the average distance between these zero crossing points is calculated. This average distance is the average period size of the signal, and thus the average frequency of the signal contained within this particular time slice can be calculated by dividing the sample rate by this average period. At step 200, the frequency of the signal and the average period size is returned to the calling function "Evaluate Filtered Data." Processing then continues at step 184 in FIG. 6B.

Referring again to that figure, once the frequency of the signal has been determined, at step 186 it is determined whether that frequency falls within the cutoff frequencies of the current filter band. If so, step 188 is executed, wherein the frequency and the amplitude is stored in the "ffreq" and the "ampi" arrays of the time slice's corresponding Higgins data structure. If the frequency does not fall within the cutoff frequencies of the current filter band, then the frequency is discarded and step 190 is executed, thereby causing the DSP 36 to return to the calling function "Do Filter Pass." Processing then continues at step 176 in FIG. 6A.

As is shown in FIG. 6A, once the "Evaluate Filter" Function has been performed, and the frequency and amplitude of the current frequency band has been determined, the DSP 36 proceeds next to program step 178. That step checks whether the last time slice has been processed. If not, then the program continues in the loop, and proceeds to program step 176 to again operate the current band filter on the next time slice, as previously described. If the last time slice has been filtered, then step 180 is performed and the processor 36 returns to the "Decompose a Speech Signal" function where processing continues at step 158 in FIG. 6.

With continued reference to FIG. 6, the processor determines at step 159 if the first filter band has just been used for this time segment. If so, the next step in the process is shown at program step 162. There, a function referred to as "Get Fundamental Frequency" is performed, which is shown in further detail in FIG. 6D, and to which reference is now made.

Beginning at step 202 of that function, the data associated with the current time segment is received. Next, the processor 36 proceeds to program step 204 and identifies, by querying the contents of the respective "ffreq" array locations, which of the time slices have frequency components that are less than 350 Hz. This range of frequencies (0 through 350 Hz) was chosen because the fundamental frequency for most speakers falls somewhere within the range of 70 to 350 Hz. Limiting the search to this range insures that only fundamental frequencies will be located. When a time slice is located that does have a frequency that falls within this range, it is placed in a histogram type data structure. The histogram is broken up into "bins," which correspond to 50 hz blocks within the 0 to 350 Hz range.

Once this histogram has been built, the DSP 36 proceeds to step 206, and determines which bin in the histogram has the greatest number of frequencies located therein. The frequencies contained within that particular bin are then averaged, and the result is the Average Fundamental Frequency (Fo) for this particular time segment. This value is then stored in DSP data memory 42.

At step 208, the DSP 36 calculates the "moving" average of the average fundamental frequency, which is calculated to be equal to the average of the Fo 's calculated for the previous time segments. In the preferred embodiment, this moving average is calculated by keeping a running average of the previous eight time segment average fundamental frequencies, which corresponds to about two seconds of speech. This moving average can be used by the processor 36 to monitor trends in the speaker's voice, such as a change in volume, and pitch, or even a change in speaker.

Once the average fundamental frequency for the time segment and the moving average of the fundamental frequency has been calculated, the processor 36 then enters a loop to determine whether the individual time slices that make up the current time segment have a fundamental frequency fo component. This determination is made at step 210, wherein the processor 36, beginning with the first time slice, compares the time slice's various frequency components (previously identified and stored within the ffreq array in the corresponding data structure) to the average fundamental frequency Fo identified in step 206. If one of the frequencies is within about 30% of that value, then that frequency is deemed to be a fundamental frequency of the time slice, and it is stored as a fundamental fo in the time slice Higgins data structure, as is indicated at program step 214. As is shown at step 212, this comparison is done for each time slice. At step 216, after each time slice has been checked, the DSP 36 returns to the Decompose a Speech Signal routine, and continues processing at step 162 in FIG. 6.

At step 160 in that figure, the processor 36 checks if the last pair of cutoff frequencies (fL and fH) has yet been used. If not, the processor 36 continues the loop at step 154, and obtains the next set of cutoff frequencies for the next filter band. The DSP 36 then continues. the filtering process as described above until the last of the filter bands has been used to filter each time slice. Thus, each time segment will be filtered at each of the filter bands. When complete, the Higgins data structure for each time slice will have been updated with each a clear identification of the frequency, and its amplitude, contained within each of the various filter bands. Advantageously, the frequency data has thus far been obtained without utilizing an FFT approach, and the problems associated with that tool have thus been avoided.

Once the final pair of cutoff frequencies has been used at step 160, step 166 causes the DSP 36 to execute a return to the main program illustrated in FIG. 4. Having completed the Decompose a Speech Signal portion of the program method, there exists a Higgins Data structure for each time slice. Contained within that structure are various sound characteristics culled from both time domain data and frequency domain data. These characteristics can now be utilized to identify the particular sound, or phoneme, carried by the signal. In the preferred embodiment, the series of program steps used to implement this portion of the program method are stored within the host program memory 58, and are executed by the Host Sound Processor 54.

This first function performed by the host sound processor 54 is illustrated in the block labeled "Point of Maximum Intelligence" shown at item 106 in FIG. 4. In this function, the processor 54 evaluates which of the Higgins data structures are critical to the identification of the phoneme sounds contained within the time segment. This reduces the amount of processing needed to identify a phoneme, and insures that phonemes are accurately identified.

One example of the detailed program steps used to implement this function are shown in FIG. 7, to which reference is now made. The process begins at step 230, where the host sound processor 54 receives each of the Higgins Data Structures for the current time segment via the host interface 52, and stores them within host data memory 60. At step 232, for all time slices containing a voiced sound, the absolute value of the slope of each filtered signal frequency is calculated, and then summed. The slope of a particular filtered signal is preferably calculated with reference to the frequencies of the signals located in the immediately adjacent time slices. Thus, for the filtered signal associated with the second frequency band, its slope is calculated by referencing its frequency with the corresponding filter signal frequencies in adjacent time slices (which are located in the second array location of the respective ffreq array). The sum of the absolute value of each filtered signal's slope for a time slice is then stored in the sumSlope variable of each applicable Higgins data structure.

The host processor 54 then proceeds to program step 234. At this step, a search is conducted for those time slices which have a sumSlope value going through a minimum and which also have an average amplitude LS that goes through a maximum. The time slices which satisfy both of these criteria are time slices where the formant frequencies are changing the least (i.e., minimum slope) and where the sound is at it highest average amplitude (i.e., highest LS), and are thus determined to be the point at which the dynamic sound has most closely reached a static or target sound. Those time slices that satisfy both criteria are identified as "points of maximum intelligence," and the corresponding PMI variable within the Higgins data structure is filled with a PMI value. Other time slices contain frequency components that are merely leading up to this target sound, and thus contain information that is less relevant to the identification of the particular phoneme.

Having identified which "voiced" time slices should be considered "points of maximum intelligence," the same is done for all time slices containing an "unvoiced" sound. This is accomplished at step 236, where each unvoiced time slice having an average amplitude LS that goes through a maximum is identified as a "point of maximum intelligence." Again, the corresponding PMI variable within the appropriate Higgins data structure is filled with a PMI value.

The host processor 54 then proceeds to program step 238 wherein the "duration" of each time slice identified as a PMI point is determined by calculating the number of time slices that have occurred since the last PMI time slice occurred. This duration value is the actual PMI value that is placed within each time slice data structure that has been identified as being a "point of maximum intelligence." The host processor 54 then returns, as is indicated at step 240, to the main calling routine shown in FIG. 4.

Referring again to that figure, the next functional block performed is the "Evaluate" function, shown at 108. This function analyzes the sound characteristics of each of the time slices identified as points of maximum intelligence, and determines the most likely sounds that occur during these time slices. This is generally accomplished by comparing the measured sound characteristics (i.e., the contents of the Higgins structure) to a set of standard sound characteristics. The sound standards have been compiled by conducting tests on a cross-section of various individual speaker's sound patterns, identifying the characteristics of each of the sounds, and then formulating a table of standard sound characteristics for each of the forty or so phonemes which make up the given language.

Referring to FIG. 8, one example of the detailed program steps used to implement the Evaluate function are illustrated. Beginning at program step 242, each of the time slices identified as PMI points are received. At step 244, the host processor 54 executes a function referred to as "Calculate Harmonic Formant Standards."

The Calculate Harmonic Formant Standards function operates on the premise that the location of frequencies within any particular sound can be represented in terms of "half-steps." The term half-steps is typically used in the musical context, but it is also a helpful in the analysis of sounds. On a musical or chromatic scale, the frequency of the notes doubles every octave. Since there are twelve notes within an octave, the frequency of two notes are related by the formula:

UPPER NOTE=(LOWER NOTE)*2n/12,

where n is the number of half-steps.

Given two frequencies (or notes), the number of half-steps between them is given by the equation: ##EQU2##

Thus, the various frequencies within a particular sound can be thought of in terms of a musical scale by calculating the distance between each component frequency and the fundamental frequency in terms of half-steps. This notion is important because it has been found that for any given sound, the distance (i.e. the number of half-steps) between the fundamental frequency and the other component frequencies of the sound are very similar for all speakers--men, women and children.

The Calculate Harmonic Formant Standards function makes use of this phenomania by building a "standard" musical table for all sounds. Specifically, this table includes the relative location of each of the sound's frequency components in terms of their distance from a fundamental frequency, wherein the distance is designated as a number of half-steps. This is done for each phoneme sound. This standard musical table is derived from the signal characteristics that are present in each sound type (phoneme), which are obtained via sample data taken from a cross-section of speakers.

Specifically, voice samples were taken from a representative group of speakers whose fundamental frequencies cover a range of about 70 Hz to about 350 Hz. The voice samples are specifically chosen so that they include all of the forty or so phoneme sounds that make up the English language. Next, the time domain signal for each phoneme sound is 2 evaluated, and all of the frequency components are extracted in the manner previously described in the Decompose function using the same frequency bands. Similarly, the amplitudes for each frequency component are also measured. From this data, the number of half steps between the particular phoneme sound's fundamental frequency and each of the sound's component frequencies is determined. This is done for all phoneme sound types. A separate x-y plot can then be prepared for each of the frequency bands for each sound. Each speaker's sample points are plotted, with the speaker's fundamental frequency (in half-steps) on the x-axis, and the distance between the measured band frequency and the fundamental frequency (in half-steps) on the y-axis. A linear regression is then performed on the resulting dam, and a resulting "best fit line" drawn through the data points. An example of such a plot is shown in FIGS. 12A-12C, which illustrates the representative data points for the sound "Ah" (PASCII sound 024), for the first three frequency bands (shown as B1, B2 and B3).

Graphs of this type are prepared for all of the phoneme sound types, and the slope and the y-intercept equations for each frequency band for each sound are derived. The results are placed in a tabular format, one preferred example of which is shown in TABLE II in Appendix A. As is shown, this table contains a phoneme sound (indicated as a PASCII value) and, for each of the bandpass frequencies, the slope (m) and the y-intercept (b) of the resulting linear regression line. Also included in the table is the mean of the signal amplitudes for all speakers, divided by the corresponding LS value, at each particular frequency band. Alternatively, the median amplitude value may be used instead.

As can be seen from the graph in FIGS. 12A-12C, the data points for each of the speakers in the test group are tightly grouped about the regression line, regardless of the speaker's fundamental frequency. This same pattern exists for most all other sounds as well. Further, the pattern extends to speakers other than those used to generate the sample data. In fact, if the fundamental frequency and the frequency band locations (in half-steps) are known for any given sound generated by any given user, the corresponding sound type (phoneme) can be determined by comparison to these standard values.

The Calculate Harmonic Formant Standards function utilizes this standard sound equations data (TABLE II) to build a representative musical table containing the standard half-step distances for each sound. Importantly, it builds this standards table so that it is correlated to a specific fundamental frequency, and specifically, it uses the fundamental frequency of the time slice currently being evaluated. The function also builds a musical table for the current time slice's measured data (i.e., the Higgins structure fo and ffreq data). The time slice "measured" data is then compared to the sound "standard" data, and the closest match indicates the likely sound type (phoneme). Since what is being compared is essentially the relative half-step distances between the various frequency components and the fundamental frequency--which for any given sound are consistent for every speaker--the technique insures that the sound is recognized independently of the particular speaker.

One example of the detailed program steps used to accomplish the "Calculate Harmonic Formant Standards" function is shown in FIG. 8A, to which reference is now made. Beginning at program step 280, the Higgins structure for the current time slice is received. Step 282 then converts that time slice into a musical scale. This is done by calculating the number half-steps each frequency component (identified in the "Decompose" function and stored in the ffreq array) is located from the fundamental frequency. These distances are calculated with the following equation: ##EQU3## where N=1 through 15, corresponding to each of the different frequencies calculated in the Decompose function and stored in the ffreq array for this time slice; and fo =the fundamental frequency for this time slice, also stored in the Higgins data structure. The value 60 is used to normalize the number of half-steps to an approximate maximum number of half-steps that occur.

The results of the calculation are stored by the host processor 54 as an array in the host processor data memory 60.

Having converted the time slice to the musical scale, the processor 54 next enters a loop to begin building the corresponding sound standards table, so it too is represented in the musical scale. Again, this is accomplished with the standard equations data (TABLE II), which is also stored as an array in host data memory 60.

Beginning at step 284, the host processor 54 obtains the standard equations data for a sound, and queries whether the current time slice contains a voiced sound. If not, the processor 54 proceeds to program step 290, where it calculates the number of half-steps each frequency component (for each of the frequency bands previously identified) is located from the fundamental frequency. The new "standards" are calculated relative to the fundamental frequency of the current time slice. The formula used to calculate these distance is: ##EQU4## where m=the slope of the standard equation line previously identified; b=the y-intercept of the standard equation line previously identified; fo =fundamental frequency of the current time slice; and the value 60 is used to normalize the number of half-steps to an approximate maximum number of half-steps that occur.

This calculation is completed for all 15 of the frequency bands. Note that unvoiced sounds do not have a "fundamental" frequency stored in the data structure's fo variable. For purposes of program step 290, the frequency value identified in the first frequency band (i.e. contained in the first location of the ffreq array) is used as a "fundamental."

If at step 284 it is determined that the current time slice is voiced, the host sound processor 54 proceeds to program step 286 and queries whether the current standard sound is a fricative. If it is a fricative sound, then the processor 54 proceeds to step 290 to calculate the standards for all of the frequency bands (one through fifteen) in the manner described above.

If the current sound is not a fricative, the host processor 54 proceeds to step 288. At that step, the standards are calculated in the same manner as step 290, but only for the frequency bands 1 through 11.

After the completion of program step 288 or step 290, the processor 54 proceeds to step 292, where it queries whether the final standard sound in the table has been processed for this time slice. If not, the next sound and its associated slope and intercept data are obtained, and the loop beginning at step 284 is re-executed. If no sounds remain, then the new table of standard values, expressed in terms of the musical scale, is complete for the current time slice (which has also been converted to the musical scale). The host processor 54 exits the routine at step 294, and returns to the Evaluate function at step 244 in FIG. 8.

Referring again to that figure, the host processor 54 next executes program step 250 to query whether the current time slice is voiced. If not, the processor 54 executes program step 246, which executes a function referred to as "Multivariate Pattern Recognition." This function merely compares "standard" sound data with "measured" time slice data, and evaluates how closely the two sets of data correspond. In the preferred embodiment, the function is used to compare the frequency (expressed in half-steps) and amplitude components of each of the standard sounds to the frequency (also expressed in half-steps) and amplitude components of the current time slice. A close match indicates that the time slice contains that particular sound (phoneme).

One example of the currently preferred set of program steps used to implement the "Multivariate Pattern Recognition" function is shown in the program flow chart of FIG. 8B, to which reference is now made. Beginning at step 260, an array containing the standard sound frequency component locations and their respective amplitudes, and an array containing the current time slice frequency component locations and their respective amplitudes, are received. Note that the frequency locations are expressed in terms of half-step distances from a fundamental frequency, calculated in the "Calculate Harmonic Formant Standards" function. The standard amplitude values are obtained from the test data previously described, examples of which are shoWn in TABLE II, and the amplitude components for each time slice are contained in the Higgins structure "amplitude" array, as previously described.

At step 262, the first sound standard contained in the standards array is compared to the corresponding time slice data. Specifically, each time slice frequency and amplitude "data point" is compared to each of the current sound standard frequency and amplitude "data points." The data points that match the closest are then determined.

Next, at program step 264, for the data points that match most closely, the Euclidean distance between the time slice data and the corresponding standard data is calculated. The Euclidean distance (ED) is calculated with the following equation: ##EQU5##

Where n=the number of data points compared; "f" indicates frequency; and "a" indicates amplitude.

At program step 266, this distance is compared to the distances found for other sound standards. If it is one of the five smallest found thus far, the corresponding standard sound is saved in the Higgins structure in the POSSIBLE PHONEMES array at step 268. The processor then proceeds to step 270 to check if this was the last sound standard within the array and, if not, the next standard is obtained at program step 272. The same comparison loop is then performed for the next standard sound. If at step 266 it is found that the calculated Euclidean distance is not one of the five smallest distances already found, then the processor 54 discards that sound as a possibility, and proceeds to step 270 to check if this was the final standard sound within the array. If not, the next sound standard is obtained at program step 272, and the comparison loop is re-executed.

This loop continues to compare the current time slice data to standard sound data until it is determined at step 270 that them are no remaining sound standards for this particular time slice. At that point, step 274 is performed, where each of the sound possibilities previously identified (up to five) are prioritized in descending order of probability. The prioritization is based on the following equation: ##EQU6##

where ED=Euclidean Distance calculated for this sound; SM=the slum of all EDs of identified sound possibilities.

The higher the probability value, the more likely that the corresponding sound is the sound contained within the time slice. Once the probabilities for each possible sound have been determined, the processor 54 proceeds to step 276, and returns to the calling routine Evaluate at step 246 in FIG. 8. The Higgins structure now contains an array of the most probable phonemes (up to five) corresponding to this particular time slice. Host Processor 54 then performs step 248 to determine if there is another time slice to evaluate. If there is, the processor 54 reenters the loop at step 242 to obtain the next time slice and continue processing. If no time slices remain, the processor 54 executes step 260 and returns to the main calling routine in FIG. 4.

If at step 250, it was instead determined that the current time slice contained a voiced sound, then the host sound processor 54 proceeds to program step 252. At this step, the host processor 54 determines whether the sound carried in the time slice is a voiced fricative, or if it is another type of voiced sound. This determination is made by inspecting the Relative Amplitude (RA) value and the frequency values contained in the ffreq array. If RA is relatively low, which in the preferred embodiment is any value less than about 65, and if there are any frequency components that are relatively high, which in the preferred embodiment is any frequency above about 6 kHz, then the sound is deemed a voiced fricative, and host 54 proceeds to program step 254. Otherwise, 54 proceeds to program step 256.

Program steps 254 and 256 both invoke the "Multivariate Pattern Recognition" routine, and both return a Higgins structure containing up to five possible sounds, as previously described. After completing program step 254, the host processor 54 will get the next time slice, as is indicated at step 248.

However, when program step 258 is completed, the host processor 54 will execute program step 258, which corresponds to a function referred to as "Adjust for Relative Amplitude." This function assigns new probability levels to each of the possible sounds previously identified by the "Multivariate Pattern Recognition" routine and stored in the Higgins data structure. This adjustment in probability is based on yet another comparison between the time slice data and standard sound data. One example of the presently preferred program steps needed to implement this function is shown in FIG. 8C, to which reference is now made.

Beginning at program step 300, the relative amplitude (RA) for the time slice is calculated using the following formula: ##EQU7## where LS is the absolute average of the amplitude for this time slice stored in the Higgins Structure; and MaxAmpl is the "moving average" over the previous 2 seconds of the maximum LS for each time segment (10,240 data points) of data.

The host processor 54 then proceeds to program step 304 and calculates the difference between the standard relative amplitude calculated in step 300, and the standard relative amplitude for each of the probable sounds contained in the Higgins data structure. The standard amplitude data is comprised of average amplitudes obtained from a representative cross-sample of speakers, an example of which is shown in TABLE III in the appendix.

Next, at program step 306 the differences are ranked, with the smallest difference having the largest rank, and the largest difference having the smallest rank of one. Proceeding next to program step 308, new probability values for each of the probable sounds are calculated by averaging the previous confidence level with the new percent rank calculated in step 306. At program step 310, the probable sounds are then re-sorted, from most probable to least probable, based on the new confidence values calculated in step 308. At step 312, the host processor 54 returns to the calling routine "Evaluate" at program step 258 in FIG. 8.

Referring again to FIG. 8 having completed the Adjust for Relative Amplitude routine, the host sound processor proceeds to program step 248 and determines whether another time slice remains. If so, the processor 54 reenters the loop at step 242, and processes a new time slice in the same manner as described above. If not, the processor 54 executes step 260 and returns to the main calling routine in FIG. 4.

The next step performed by the sound recognition host processor 54 is shown at block 110 in FIG. 4 and is referred to as the "Compress Phones" function. As already discussed, this function discards those time slices in the current time segment that are not designated "points of maximum intelligence." In addition, it combines any contiguous time slices that represent "quiet" sounds. By eliminating the unnecessary time slices, all that remains are the time slices (and associated Higgins structure data) needed to identify the phonemes contained within the current time segment. This step further reduces overall processing requirements and insures that the system is capable of performing sound recognition in substantially real time.

One presently preferred example of the detailed program steps used to implement the "Compress Phones" function is shown in FIG. 9, to which reference is now made. Beginning at program step 316, the host sound processor 54 receives the existing sequence of time slices and the associated Higgins data structures. At program step 318, processor 54 eliminates all Higgins structures that do not contain PMI points. Next, at program step 320 the processor 54 identifies contiguous data structures containing "quiet" sections, and reduces those contiguous sections into a single representative data structure. The PMI duration value in that single data structure is incremented so as to represent all of the contiguous "quiet" structures that were combined.

At this point, there exists in the host processor data memory 60 a continuous stream of Higgins data structures, each of which contains sound characteristic data and the possible phoneme(s) associated therewith. All unnecessary, irrelevant and/or redundant aspects of the time segment have been discarded so that the remaining data stream represents the "essence" of the incoming speech signal. Importantly, these essential characteristics have been culled from the speech signal in a manner that is not dependent on any one particular speaker. Further, they have been extracted in a manner such that the speech signal can be processed in substantially real time--that is, the input can be received and processed at normal rate of speech.

Having reduced the Higgins structure data, the Compress Phones function causes the sound recognition host processor 54 to place that data in host data memory 60 in program step 324. Proceeding next to program step 326, the host sound processor 54 returns to the main portion of the program method in FIG. 4.

As is shown in that figure, the next portion of the program method corresponds to the function referred to as the "Linguistic Processor." The Linguistic Processor is that portion of the method which further analyzes the Higgins structure data and, by applying a series of higher level linguistic processing techniques, identifies the word or phrase that is contained within the current time segment portion of the incoming speech signal.

Although alterative linguistic processing techniques and approaches could be used, one presently preferred set of program steps used to implement the Linguistic Processor is shown in the flow chart of FIG. 10. Beginning at program step 350 of that function, the host sound processor 54 receives the set of Higgins structure data created by the previously executed Compress Phones function. As already discussed, this data represents a stream of the possible phonemes contained in the current time segment portion of the incoming speech signal. At program step 352, the processor 54 passes this data to a function referred to as "Dictionary Lookup."

In one preferred embodiment, the Dictionary Lookup function utilizes a phonetic-English dictionary that contains the English spelling of a word along with its corresponding phonetic representation. The dictionary can thus be used to identify the English word that corresponds to a particular stream of phonemes. The dictionary is stored in a suitable database structured format, and is placed within the dictionary portion of computer memory 62. The phonetic dictionary can be logically separated into several separate dictionaries. For instance, in the preferred embodiment, the first dictionary contains a database of the most commonly used English words. Another dictionary may include a database that contains a more comprehensive Webster-like collection of words. Other dictionaries may be comprised of more specialized words, and may vary depending on the particular application. For instance, there may be a user defined dictionary, a medical dictionary, a legal dictionary, and so on.

All languages can be described in terms of a particular set of phonetic sounds. Thus, it will be appreciated that although the preferred embodiment utilizes an English word dictionary, any other phonetic to non-English language dictionary could be used.

Basically, Dictionary Lookup scans the appropriate dictionary to determine if the incoming sequence of sounds (as identified by the Higgins data structures) form a complete word, or the beginnings of a possible word. To do so, the sounds are placed into paths or "sequences" to help detect, by way of the phonetic dictionary, the beginning or end of possible words. Thus, as each phoneme sound is received, it is added to the end of a all non-completed "sequences." Each sequence is compared to the contents of the dictionary to determine if it leads to a possible word. When a valid word (or set of possible words) is identified, it is passed to the next functional block within the Linguistic Processor portion of the program for further analysis.

By way of example and not limitation, FIG. 10A illustrates one presently preferred set of program steps used to implement the Dictionary Lookup function. The function begins at program step 380, where it receives the current set of Higgins structures corresponding to the current time segment of speech. At program step 384, the host sound processor 54 obtains a phoneme sound (as represented in a Higgins structure) and proceeds to program step 386 where it positions a search pointer within the current dictionary that corresponds to the first active sequence. An "active" sequence is a sequence that could potentially form a word with the addition of a new sound or sounds. In contrast, a sequence is deemed "inactive" when it is determined that there is no possibility of forming a word with the addition of new sounds.

Thus, at program step 386 the new phonetic sound is appended to the first active sequence. At program step 388, the host processor 54 checks, by scanning the current dictionary contents, whether the current sequence either forms a word, or whether it could potentially form a word by appending another sound(s) to it. If so, the sequence is updated by appending to it the new phonetic sound at program step 390. Next, at program step 392, the host processor determines whether the current sequence forms a valid word. If it does, a `new sequence` flag is set at program step 394, which indicates that a new sequence should be formed beginning with the very next sound. If a valid word is not yet formed, the processor 54 skips step 394, and proceeds directly to program step 396.

If at step 388 the host processor 54 instead determines, after scanning the dictionary database, that the current sequence would not ever lead to a valid word, even if additional sounds were appended, then the processor 54 proceeds to program step 398. At this step, this sequence is marked "inactive." The processor 54 then proceeds to program step 396.

At step 396, the processor 54 checks if there are any more active sequences to which the current sound should be appended. If so, the processor 54 will proceed to program step 400 and append the sound to this next active sequence. The processor 54 will then re-execute program step 388, and process this newly formed sequence in the same manner described above.

If at program step 396 it is instead determined that there are no remaining active sequences, then host sound processor 54 proceeds to program step 402. There, the `new sequence` flag is queried to determine if it was set at program step 394, thereby indicating that the previous sound had created a valid word in combination with an active sequence. If set, the processor will proceed to program step 406 and create a new sequence, and then go to program step 408. If not set, the processor 54 will instead proceed to step 404, where it will determine whether all sequences are now inactive. If they are, processor 54 will proceed immediately to program step 408, and if not, the processor 54 will instead proceed to step 406 where it will open a new sequence before proceeding to program step 408.

At program step 408, the host sound processor 54 evaluates whether a primary word has been completed, by querying whether all of the inactive sequences, and the first active sequence result in a common word break. If yes, the processor 54 will output all of the valid words that have been identified thus far to the main calling routine portion of the Linguistic Processor. The processor 54 will then discard all of the inactive sequences, and proceed to step 384 to obtain the next Higgins structure sound. If at step 408 it is instead determined that a primary word has not yet been finished, the processor 54 will proceed directly to program step 384 to obtain the next Higgins structure sound. Once a new sound is obtained at step 384, the host processor 54 proceeds directly to step 386 and continues the above described process.

As the Dictionary Lookup function extracts words from the Higgins structure data, there may certain word possibilities that have not yet been resolved. Thus, the Linguistic Processor may optionally include additional functions which further resolve the remaining word possibilities. One such optional function is referred to as the "Word Collocations" function, shown at block 354 in FIG. 10.

Generally, the Word Collocations function monitors the word possibilities that have been identified by the Dictionary Lookup function to see if they form a "common" word collocation. A set of these common word collocations are stored in a separate dictionary database within dictionary memory 64. In this way, certain word possibilities can be eliminated, or at least assigned lower confidence levels, because they do not fit within what is otherwise considered a common word collocation. One presently preferred example of the program steps used to implement this particular function are shown, by way of example and not limitation, in FIG. 10B, to which reference is now made.

Beginning at program step 420, a set of word possibilities are received. Beginning with one of those words at step 422, the host sound processor 54 next proceeds to program step 424 where it obtains any collocation(s) that have been formed by preceding words. The existence of such collocations would be determined by continuously comparing words and phrases to the collocation dictionary contents. If such a collocation or collocations exist, then the current word possibility is tested to see if it fits within the collocation context. At step 428, those collocations which no longer apply are discarded. The processor 54 then proceeds to step 430 to determine if any word possibilities remain, and if so, the remaining word(s) is also tested within the collocation context beginning at program step 422.

Once this process has been applied to all word possibilities, the processor 54 identifies which word, or words, were found to "fit" within the collocation, before returning, via program step 436, to the main Linguistic Processor routine. Based on the results of the Collocation routine, certain of the remaining word possibilities can then be eliminated, or at least assigned a lower confidence level.

Another optional function that can be used to resolve remaining word possibilities is the "Grammar Check" function, shown at block 356 in FIG. 10. This function evaluates a word possibility by applying certain grammatical rules, and then determining whether the word complies with those rules. Words that do not grammatically fit can be eliminated as possibilities, or assigned lower confidence levels.

By way of example, the Grammar Check function can be implemented with the program steps that are shown in FIG. 10C. Thus, at step 440, a current word possibility along with a preceding word and a following word are received. Then at step 442, a set of grammar rules, stored in a portion of host sound processor memory, are queried to determine what "part of speech" would best fit in the grammatical context of the preceding word and the following word. If the current word possibility matches this "part of speech" at step 444, then that word is assigned a higher confidence level before returning to the Linguistic Processor at step 446. If the current word does not comply with the grammatical "best fit" at step 444, then it is assigned a low confidence level and returned to the main routine at step 446. Again, this confidence level can then be used to further eliminate remaining word possibilities.

Referring again to FIG. 10, having completed the various functions which identify the word content of the incoming speech signal, the Linguistic Processor function causes the host sound processor 54 to determine the number of word possibilities that still exist for any given series of Higgins structures.

If no word possibilities have yet been identified, then the processor 54 will determine, at program step 366, if there remains a phonetic dictionary database (i.e., a specialized dictionary, a user defined dictionary, etc.) that has not yet been searched. If so, the processor 54 will obtain the new dictionary at step 368, and then re-execute the searching algorithm beginning at program step 352. If however no dictionaries remain, then the corresponding unidentified series of phoneme sounds (the unidentified "word") will be sent directly to the Command Processor portion of the program method, which resides on Host computer 22.

If at program step 358 more than one word possibility still remains, the remaining words are all sent to the Command Processor. Similarly, if only one word possibility remains, that word is sent to the directly to the Command Processor portion of the algorithm. Having output the word, or possible words, program step 370 causes the host sound processor 54 to return to the main algorithm, shown on FIG. 4.

As words are extracted from the incoming speech signal by the Linguistic Processor, they are immediately passed to the next function in the overall program method referred to as the "Command Processor," shown at function block 114 in FIG. 4. In the preferred embodiment, the Command Processor is a series of program steps that are executed by a Host Computer 22, such as a standard desktop personal computer. As already noted, the host computer 22 receives the incoming words by way of a suitable communications medium, such as a standard RS-232 cable 24 and interface 66. The Command Processor then receives each word, and determines the manner by which it should be used on the host computer 22. For example, a spoken word may be input as text directly into an application, such as a wordprocessor document. Conversely, the spoken word may be passed as a command to the operating system or application.

Referring next to FIG. 11, illustrated is one preferred example of the program steps used to implement the Command Processor function. To begin, program step 450 causes the host computer 22 to receive a word created by the Linguistic Processor portion of the algorithm. The host computer 22 then determines, at step 452, whether the word received is an operating system command. This is done by comparing the word to the contents of a definition file database, which defines all words that constitute operating system commands. If such a command word is received, it is passed directly to the host computer 22 operating system, as is shown at program step 454.

If the incoming word does not constitute an operating system command, step 456 is executed, where it is determined if the word is instead an application command, as for instance a command to a wordprocessor or spreadsheet. Again, this determination is made by comparing the word to another definition file database, which defines all words that constitute an application command. If the word is an application command word, then it is passed directly, at step 458, to the intended application.

If the incoming word is neither a operating system command, or an application command, then program step 460 is executed, where it is determined whether the Command Processor is still in a "command mode." If so, the word is discarded at step 464, and essentially ignored. However, if the Command Processor is not in a command mode, then the word will be sent directly to the current application as text.

Once a word is passed as a command to either the operating system or application at program steps 454 and 458, the host computer 22 proceeds to program step 466 to determine whether the particular command sequence is yet complete. If not, the algorithm remains in a "command mode," and continues to monitor incoming words so as to pass them as commands directly to the respective operating system or application. If the command sequence is complete at step 466, then the algorithm will exit the command mode at program step 470.

In this way, the Command Processor acts as a front-end to the operating system and/or to the applications that are executing on the host computer 22. As each new word is received, it is selectively directed to the appropriate computer resource. Operating in this manner, the system and method of the current invention act as a means for entering data and/or commands to a standard personal computer. As such, the system essentially replaces, or supplements other computer input devices, such as keyboards and pointing devices.

Attached hereto at Appendix B, and incorporated herein by reference, is an example of a computer program listing written in the "C" programming language, which serves to illustrate one way in which the method of the present invention was implemented to perform real-time, user-independent speech recognition. It should be recognized that the system and method of the present invention are not intended to be limited by the program listing contained in Appendix B, which is merely an illustrative example, and that the method could be implemented using virtually any other programming language other than "C."

III. SUMMARY AND SCOPE OF THE INVENTION

In summary, the system and method of the present invention for speech recognition provides a powerful and much needed tool for providing user independent speech recognition. Importantly, the system and method extracts only the essential components of an incoming speech signal. The system then isolates those components in a manner such that the underlying sound characteristics that are common to all speakers can be identified, and thereby used to accurately identify the phonetic make-up of the speech signal. This permits the system and method to recognize speech utterances from any speaker of a given language, without requiring the user to first "train" the system with specific voice characteristics.

Further, the system and method implements this user independent speech recognition in a manner such that it occurs in substantially "real time." As such, the user can speak at normal conversational speeds, and is not required to pause between each word.

Finally, the system utilizes various linguistic processing techniques to translate the identified phonetic sounds into a corresponding word or phrase, of any given language. Once the phonetic stream is identified, the system is capable of recognizing a large vocabulary of words and phrases.

While the system and method of the present invention has been described in the context of the presently preferred embodiment and the examples illustrated and described herein, the invention may be embodied in other specific ways or in other specific forms without departing from its spirit or essential characteristics. Therefore, the described embodiments and examples are to be considered in all respects only as illustrative and not restrictive. The scope of the invention is, therefore, indicated by the appended claims rather than by the foregoing description, and all changes which come within the meaning and range of equivalency of the claims are to be embraced within their scope.

              TABLE II______________________________________APPENDIX ASOUND                    Y-PASCII FREQ     SLOPE    INTERCEPTVALUE  BAND #   (m)      (b)       AMPLITUDE______________________________________ 1      1       23.925   0.0639    0.73378 1      2       43.1006  0.116964  0.08242 1      3       54.5453  0.1132    0.01025 1      4       60.7934  0.111916  0.01257 1      5       62.7092  0.0989    0.06235 1      6       66.9046  0.105248  0.07415 1      7       68.9042  0.101159  0.098 1      8       70.9394  0.102078  0.05573 1      9       73.8657  0.103871  0.0297 1     10       76.6542  0.109661  0.01606 1     11       78.566   0.105545  0.02196 1     12       0        0         0 1     13       0        0         0 1     14       0        0         0 1     15       0        0         0 3      1       31.0818  0.0948    0.7639 3      2       37.6375  0.0787    0.2279 3      3       54.8824  0.115936  0.02602 3      4       59.882   0.103487  0.05287 3      5       61.8428  0.097235  0.11788 3      6       67.4577  0.107712  0.10825 3      7       69.8282  0.106478  0.04873 3      8       71.9363  0.104027  0.02985 3      9       74.1275  0.105246  0.02271 3     10       75.6143  0.102906  0.00936 3     11       79.3344  0.106665  0.01144 3     12       0        0         0 3     13       0        0         0 3     14       0        0         0 3     15       0        0         0 7      1       35.8081  0.117513  0.7151 7      3       55.9232  0.122236  0.05651 7      4       59.2746  0.105329  0.201 7      5       64.3502  0.111596  o.15908 7      6       66.8912  0.105726  0.10852 7      7       70.5895  0.110907  0.07466 7      8       72.3561  0.108349  0.03763 7      9       74.6623  0.108032  0.02601 7     10       76.825   0.11056   0.0184 7     11       79.5416  0.110216  0.02638 7     12       0        0         0 7     13       0        0         0 7     14       0        0         0 7     15       0        0         024      1       26.3645  0.0772    0.582024      2       43.4946  0.095061  0.7198124      3       50.796   0.0974    0.4664824      4       60.6949  0.122010  0.0533224      5       65.2771  0.116403  0.0380624      6       66.9481  0.106186  0.0573524      7       71.0327  0.1114422 0.0365424      8       72.5388  0.108871  0.0303124      9       76.0082  0.116995  0.0137824     10       77.3385  0.112669  0.01724     11       78.6243  0.104959  0.0159124     12       0        0         024     13       0        0         024     14       0        0         024     15       0        0         0 9      1       27.3808  0.0891    0.6873 9      2       46.0161  0.117744  0.6969 9      3       57.4503  0.132157  0.2288 9      4       59.863   0.113996  0.3164 9      5       67.1216  0.130564  0.16726 9      6       68.5702  0.119971  0.10475 9      7       72.1892  0.122477  0.04561 9      8       73.8496  0.11908   0.04229 9      9       77.1308  0.125179  0.02519 9     10       78.0586  0.118421  0.02961 9     11       81.6235  0.125473  0.02507 9     12       0        0         0 9     13       0        0         0 9     14       0        0         0 9     15       0        0         014      1       29.9976  0.0967    0.603514      2       40.7298  0.0901    0.7317414      3       55.0417  0.117045  0.2434414      4       58.4921  0.107211  0.1090414      5       65.8377  0.119586  0.0451714      6       65.9093  0.100399  0.0518314      7       70.9514  0.113684  0.0356414      8       72.519   0.107896  0.0239814      9       75.3182  0.113199  0.0190614     10       76.5463  0.108146  0.0142514     11       79.4491  0.109836  0.0192914     12       0        0         014     13       0        0         014     14       0        0         014     15       0        0         017      1       26.9756  0.076984  0.8465617      2       51.8834  0.148419  0.132717      3       50.9061  0.0955    0.0649417      4       60.211   0.111777  0.0172217      5       63.4817  0.10496   0.0170417      6       67.1155  0.106036  0.0118717      7       70.9826  0.112958  0.010217      8       71.1014  0.0997    0.0084417      9       74.2932  0.106116  0.0049817     10       76.5634  0.107109  0.004317     11       80.2467  0.0114159 0.0032817     12       0        0         017     13       0        0         017     14       0        0         017     15       0        0         021      1       35.6987  0.118874  0.816921      2       42.9284  0.104448  0.628221      3       51.6091  0.106709  0.0995421      4       59.6202  0.108802  0.0100421      5       64.0317  0.107957  0.0151921      6       66.9097  0.10484   0.0139421      7       70.2666  0.107929  0.0166421      8       71.7338  0.102196  0.0117221     9        75.2727  0.1112    0.004221     10       76.7847  0.107923  0.0033421     11       79.5333  0.109177  0.007621     12       0        0         021     13       0        0         021     14       0        0         021     15       0        0         026      1       94.161   0.346415  0.468726      2       28.8099  0.0448    0.846626      3       55.6297  0.107713  0.0975126      4       40.9908  0.025     0.1444326      5       63.7703  0.103867  0.0884726      6       56.7514  0.0615    0.0257826      7       64.7022  0.0792    0.0234426      8       97.9576  0.22901   0.0123826      9       66.7865  0.0708    0.0042126     10       72.7685  0.087492  0.0063326     11       74.6368  0.0865    0.0062126     12       0        0         026     13       0        0         026     14       0        0         026     15       0        0         029      1       37.5589  0.13441   0.730329      2       29.1422  0.0426    0.640929      3       55.5325  0.11215   0.142129      4       56.7644  0.095904  0.1855329      5       62.0948  0.103664  0.0465829      6       66.5342  0.104791  0.0113229      7       68.3164  0.0982    0.009529      8       71.9616  0.104908  0.0117329      9       73.2931  0.100259  0.0045529     10       75.6625  0.102199  0.0050329     11       77.4381  0.0989    0.0052529     12       0        0         029     13       0        0         029     14       0        0         029     15       0        0         031      1       24.0535  0.065356  0.5902231      2       46.3754  0.123127  0.0609331      3       50.9352  0.091369  0.0410731      4       56.8214  0.0948    0.0280131      5       60.8415  0.089737  0.031931      6       63.9034  0.0906    0.0257931      7       66.7104  0.0894    0.0102231      8       69.1107  0.0879    0.0095631      9       71.9378  0.094015  0.0082731     10       73.6224  0.0913    0.0038931     11       77.2013  0.0941    0.0056231     12       0        0         031     13       0        0         031     14       0        0         031     15       0        0         033      1       36.1683  0.136196  0.638633      2       40.7677  0.0997    0.0857933      3       51.0809  0.0938    0.0194733      4       57.2837  0.0961    0.0206433      5       61.365   0.0925    0.0231433      6       64.1689  0.0924    0.0172833      7       67.4613  0.0944    0.0075433      8       66.918   0.0806    0.0040433      9       72.5547  0.0951    0.0052533     10       74.3771  0.095119  0.0026433     11       77.5436  0.0966    0.00333     12       0        0         033     13       0        0         033     14       0        0         033     15       0        0         036      1       25.128   0.0677    0.642836      2       42.9834  0.110396  0.1114436      3       50.5331  0.0918    0.0430236      4       57.1574  0.0935    0.018736      5       60.3679  0.0872    0.0372136      6       64.1232  0.0916    0.0361136      7       67.7702  0.0953    0.0165836      8       69.967   0.0934    0.01336      9       71.8082  0.0916    0.0067336     10       74.66    0.0975    0.0061436     11       77.0475  0.0955    0.007236     12       0        0         036     13       0        0         036     14       0        0         036     15       0        090      1       34.559   0.117681  0.845590      2       45.7616  0.123735  0.689790      3       52.5577  0.110983  0.0492490      4       60.4452  0.116582  0.007690      5       65.0779  0.113763  0.0087290      6       66.9828  0.107816  0.015290      7       70.2725  0.108867  0.0117890      8       72.0092  0.106249  0.0136990      9       75.4537  0.113103  0.0070590     10       76.8398  0.110225  0.0056290     11       79.5101  0.110944  0.0096190     12       0        0         090     13       0        0         090     14       0        0         090     15       0        0         094      1       33.01    0.10304   0.935394      2       19.5992  0.0222    0.689494      3       54.2337  0.102615  0.1463194      4       58.7361  0.106557  0.1275694      5       62.8017  0.106312  0.0225794      6       69.182   0.120343  0.0213594      7       70.1864  0.108033  0.0088194      8       71.856   0.105312  0.0056194      9       75.8229  0.114387  0.0019494     10       76.1835  0.10575   0.0015194     11       79.8682  0.110951  0.0018294     12       0        0         094     13       0        0         094     14       0        0         094     15       0        0         058      1       30.5155  0.104315  0.531758      2       41.1473  0.098     0.094558      3       52.6775  0.101027  0.0587558      4       57.3355  0.0976    0.0441358      5       61.881   0.0968    0.0392155      6       65.1193  0.0969    0.0326558      7       68.0574  0.0971    0.0277358      8       69.5643  0.092299  0.0209858      9       72.7544  0.0979    0.0163758     10       74.551   0.096685  0.0143358     11       77.332   0.098928  0.0199758     12       79.5717  0.095093  0.0138558     13       82.1972  0.0959    0.0126358     14       84.515   0.0962    0.0142558     15       86.3601  0.0958    0.0148660      1       29.8209  0.097751  0.584360      2       45.0992  0.117747  0.0893460      3       54.3205  0.10593   0.0859160      4       58.4529  0.10563   0.0583760      5       64.9092  0.112361  0.0436660      6       66.8778  0.107116  0.051460      7       71.3666  0.115105  0.0313460      8       72.4539  0.107843  0.0240660      9       74.8978  0.10941   0.0170660     10       77.1449  0.110471  0.0152760     11       79.5827  0.110523  0.0206160     12       82.3536  0.110665  0.0122660     13       84.722   0.108912  0.0125260     14       86.9515  0.109566  0.0105760     15       88.7968  0.10925   0.0133162      1       30.3312  0.109582  0.56662      2       39.1704  0.0809    0.0575662      3       55.2685  0.113607  0.0472362      4       58.3998  0.105309  0.0290662      5       65.0247  0.113936  0.0256362      6       66.6728  0.105247  0.0239462      7       70.1915  0.108396  0.019562      8       72.2755  0.106131  0.0209162      9       74.8135  0.10855   0.0304162     10       76.3984  0.106234  0.0362362     11       78.4797  0.103104  0.0826962     12       81.142   0.104107  0.0500862     13       84.3696  0.107884  0.0389662     14       85.96    0.10425   0.0330162     15       88.0833  0.15053   0.0250366      1       24.9269  0.0789    0.618866      2       48.3858  0.135315  0.0615566      3       54.5496  0.109284  0.0228266      4       58.3009  0.100643  0.0758366      5       64.9467  0.11328   0.103766      6       66.4737  0.103452  0.155566      7       69.2804  0.104905  0.126366      8       71.5304  0.103333  0.098166      9       73.733   0.103675  0.083966     10       76.5636  0.108794  0.0635866     11       78.8226  0.107965  0.081366     12       81.1678  0.10551   0.0316366     13       84.3501  0.108437  0.0173866     14       86.1132  0.106013  0.0116966     15       87.6284  0.103334  0.00849______________________________________

              TABLE III______________________________________RELATIVE-AMPLITUDE STANDARDS                  RELATIVEPHONEME       PASCII   AMPLITUDESOUND         VALUE    STANDARD______________________________________ah            23       95uh            22       95ah            24       95O             21       95a              9       85u             19       85er            29       85e              7       75A              5       75oo            17       75i              3       75w             85       75ee             1       75r             94       75y             82       75l             90       75sh            65       65ng            36       65ch            116      55m             31       55n             33       50si            66       50j             115      50t             41       40g             48       40k             47       40˜th     60       40z             62       40s             61       35h             76       35d             42       30v             58       30b             40       30p             39       25f             57       25th            59       20______________________________________ ##SPC1##
Patent Citations
Cited PatentFiling datePublication dateApplicantTitle
US3581192 *Nov 13, 1968May 25, 1971Hitachi LtdFrequency spectrum analyzer with displayable colored shiftable frequency spectrogram
US3703609 *Nov 23, 1970Nov 21, 1972E Systems IncNoise signal generator for a digital speech synthesizer
US3838217 *Mar 1, 1971Sep 24, 1974J DreyfusAmplitude regulator means for separating frequency variations and amplitude variations of electrical signals
US3938394 *Nov 30, 1973Feb 17, 1976Ird Mechanalysis, Inc.Combination balance analyzer and vibration spectrum analyzer
US3940565 *Jul 27, 1973Feb 24, 1976Klaus Wilhelm LindenbergTime domain speech recognition system
US3969972 *Apr 2, 1975Jul 20, 1976Bryant Robert LMusic activated chromatic roulette generator
US4181813 *May 8, 1978Jan 1, 1980John MarleySystem and method for speech recognition
US4452079 *Sep 27, 1982Jun 5, 1984Cooper Industries, Inc.Acoustic tachometer
US4658252 *Jul 7, 1986Apr 14, 1987Gte Government Systems CorporationEncoder/decoder for card entry system
US4780906 *Feb 17, 1984Oct 25, 1988Texas Instruments IncorporatedSpeaker-independent word recognition method and system based upon zero-crossing rate and energy measurement of analog speech signal
US4817154 *Dec 9, 1986Mar 28, 1989Ncr CorporationMethod and apparatus for encoding and decoding speech signal primary information
US4852170 *Dec 18, 1986Jul 25, 1989R & D AssociatesReal time computer speech recognition system
US4862503 *Jan 19, 1988Aug 29, 1989Syracuse UniversityVoice parameter extractor using oral airflow
US4975957 *Apr 24, 1989Dec 4, 1990Hitachi, Ltd.Character voice communication system
US4991216 *Mar 23, 1990Feb 5, 1991Matsushita Electric Industrial Co., Ltd.Method for speech recognition
US4998280 *Sep 19, 1989Mar 5, 1991Hitachi, Ltd.Speech recognition apparatus capable of discriminating between similar acoustic features of speech
US5027410 *Nov 10, 1988Jun 25, 1991Wisconsin Alumni Research FoundationAdaptive, programmable signal processing and filtering for hearing aids
US5065432 *Oct 31, 1989Nov 12, 1991Kabushiki Kaisha ToshibaSound effect system
US5068900 *Jul 16, 1990Nov 26, 1991Gus SearcyVoice recognition system
US5091948 *Mar 15, 1990Feb 25, 1992Nec CorporationSpeaker recognition with glottal pulse-shapes
US5121434 *Jun 14, 1989Jun 9, 1992Centre National De La Recherche ScientifiqueSpeech analyzer and synthesizer using vocal tract simulation
US5166981 *May 16, 1990Nov 24, 1992Sony CorporationAdaptive predictive coding encoder for compression of quantized digital audio signals
US5202926 *Sep 12, 1991Apr 13, 1993Oki Electric Industry Co., Ltd.Phoneme discrimination method
US5299125 *Feb 8, 1993Mar 29, 1994Semantic Compaction SystemsNatural language processing system and method for parsing a plurality of input symbol sequences into syntactically or pragmatically correct word messages
US5321608 *Nov 27, 1991Jun 14, 1994Hitachi, Ltd.Method and system for processing natural language
Non-Patent Citations
Reference
1Ackenhusen & Oh, Single-Ship Implementation of Feature Measurement for LPC-Based Recognition, AT&T Technical Journal, vol. 64, No. 8, pp. 1787-1805, Oct. 1985.
2Ackenhusen, All, Bishop, Ross & Thorkildsen, Single-Board General-Purpose Speech Recognition System, AT&T Technical Journal, vol. 65, No. 5, pp. 48-59, Sep./Oct. 1986.
3Advance in Computer Speech Recognition, Reader Service, No. 128, Mar. 1985.
4Ainsworth & Pratt, Feedback Strategies for Error Correction in Speech Recognition Systems, International Journal of Man-Machine Studies, vol. 36, pp. 833-842, 1992.
5Ainsworth, Technical Note: Theoretical and Simulation Approaches to Error Correction Strategies in Automatic Speech Recognition, International Journal of Man-Machine Studies, vol. 39, pp. 517-520, 1993.
6Akagi, Modeling of Contextual Effects Based on Spectral Peak Interaction, J. Acoustical Society of America.
7Al Fine-Tunes Speech Recognition, Electronics, pp. 24-25, May 19, 1986.
8Anastasako, Kubala, Makhoul & Schwartz, Adaptation to New Microphones Using Tied-Mix Normalization, IEEE 1994.
9Anderson, Cross & Lamb, Listening Computers Broaden Their Vocabulary, New Scientist, p. 38, Aug. 4, 1988.
10 *Anderson, Dalsgaard & Barry, On the Use of Data-Driven Clustering Technique for Identification of Poly-and Mono-phonemes for four European Languages, IEEE 1994.
11Andrews, IBM and Apple Work to Perfect Voice Input News & Views.
12 *Ariki, Mizuta, Nagata, Sakai, Spoken-Word Recognition Using Dynamic Features Analyzed by Two-Dimensional Cepstrum, IEE Proceedings, vol. 136, Pt. 1, No. 2, pp. 133 140, Apr. 1989.
13Ariki, Mizuta, Nagata, Sakai, Spoken-Word Recognition Using Dynamic Features Analyzed by Two-Dimensional Cepstrum, IEE Proceedings, vol. 136, Pt. 1, No. 2, pp. 133-140, Apr. 1989.
14 *Arslan & Hansen, Minimum Cost Based Phoneme Class Detection for Improved Iterative Speech Enhancement, IEEE 1994.
15Atal & Rabiner, Speech Research Directions, AT&T Technical Journal, vol. 65, No. 5, pp. 75-88, Sep./Oct. 1986.
16Baber & Hone, Modeling Error, Recovery and Repair in Automatic Speech Recognition, International Journal of Man-Machine Studies, vol. 39, pp. 495-515, 1993.
17Baber, Ushers, Stammers & Taylor, Feedback Requirements for Automatic Speech Recognition in the Process Control Room, International Journal of Man-Machine Studies, vol. 37, No. 6, Dec. 1992.
18 *Barnard, Cole, Vea & Alleva, Pitch Detection with a Neural-Net Classifier, IEEE Transactions on Signal Processing, vol. 39, No. 2, pp. 298 307, Feb. 1991.
19Barnard, Cole, Vea & Alleva, Pitch Detection with a Neural-Net Classifier, IEEE Transactions on Signal Processing, vol. 39, No. 2, pp. 298-307, Feb. 1991.
20Bartelt, Lohmann, & Wimitzer, Phase and Amplitude Recovery From Bispectra, Appfied Optics, vol. 23, No. 18, pp. 3121-3129, Sep. 15, 1984.
21Bengio, Cardin, DeMori & Merlo, Programmable Execution of Multi-Layered Networks for Automatic Speech Recognition, Communications of the ACM, vol. 32, No. 2, pp. 195-199, Feb. 1989.
22 *Bengio, De Mori, Flammia & Kompe, Global Optimization of Neural Network-Hidden Markov Model Hybird, IEEE Transactions on Neural Networks, vol. 3, No. 2, pp. 253 259, Mar. 1992.
23Bengio, De Mori, Flammia & Kompe, Global Optimization of Neural Network-Hidden Markov Model Hybird, IEEE Transactions on Neural Networks, vol. 3, No. 2, pp. 253-259, Mar. 1992.
24Bergh, Soong & Rabiner, Incorporation of Temporal Structure Into a Vector-Quantization-Based Processor for Speaker-Independent, Isolated-Word Recognition, AT&T Technical Journal, vol. 64, No 5, pp. 1047-1063, May/Jun. 1985.
25Biermann, Fineman & Heidlage, A Voice-and Touch-Driven Natural Language Editor and its Performance, International Journal of Man-Machine Studies, vol. 37, No. 1, pp. 1-21, July 1992.
26Biermann, Rodman, Rubin & Heidlage, Natural Language with Discrete Speech as a Mode for Human-to-Machine Communication, Communications of the ACM, vol. 28, No. 6, pp. 628-636, Jun. 1985.
27Black, An Experiment in Computational Discrimination of English Word Senses,IBM J. Res Develp, vol. 32, No. 2, pp. 185-194, Mar. 1988.
28Bloothooft & Plomp, The Sound Level of the Singer's Formant in Professional Singing, J. Acoustical Societ of America, vol. 79, No. 6, pp. 2028-2033, Jun. 1986.
29Borg, A Broad-Band Amplitude-Independent Phase Measuring System, J. Phys. Sci. Instrum, vol. 20, pp. 1216-1220, 1987.
30 *Bourlard, D hoore & Boite, Optimizing Recognition and Rejections Performance in Workspotting Systems, IEEE 1994.
31Bourlard, D'hoore & Boite, Optimizing Recognition and Rejections Performance in Workspotting Systems, IEEE 1994.
32Bourlard, Kamp, Ney & Wellekens, Speaker-Dependent Connected Speech Recognition via Dynamic Programming and Statistical Methods, Bibliotheca Phonetica, No. 12, pp. 115-148, 1985.
33 *Brieseman, Thorpe & Bates, Nontactile Estimation of Glottal Excitation Characteristics of Voiced Speech, IEEE Proceedings, vol. 134, Pt. A, No. 10, pp. 807 813, Dec. 1987.
34Brieseman, Thorpe & Bates, Nontactile Estimation of Glottal Excitation Characteristics of Voiced Speech, IEEE Proceedings, vol. 134, Pt. A, No. 10, pp. 807-813, Dec. 1987.
35Bringham, The Fast Fourier Transform and its Applications, Prentice Hall Signal Processing Series.
36Bristow, Electronic Speech Recognition, ISBN, 1986.
37 *Brown, McGee, Rabiner & Wilpon, Training Set Design for Connected Speech Recognition, IEEE Transactions on Signal Processing, vol. 39, No. 6, pp. 1268 1281, Jun. l991.
38Brown, McGee, Rabiner & Wilpon, Training Set Design for Connected Speech Recognition, IEEE Transactions on Signal Processing, vol. 39, No. 6, pp. 1268-1281, Jun. l991.
39Business System Recognizes Spoken English Sentences, Computer, pp. 94-95, Jan, 1985.
40Candy, O'Brien & Edmonds, End-User Manipulation of a Knowledge-Based System: A Study of An Expert's Practice, International Journal of Man-Machine Studies, vol. 38, pp. 129-145, 1993.
41 *Casacuberta, Some Relations Among Stochastic Finite State Networks Used in Automatic Speech Recognition, IEEE Transactions on Pattern Analysis & Machine Intelligence, vol. 12, No. 7, pp. 691 704, Jul. 1990.
42Casacuberta, Some Relations Among Stochastic Finite State Networks Used in Automatic Speech Recognition, IEEE Transactions on Pattern Analysis & Machine Intelligence, vol. 12, No. 7, pp. 691-704, Jul. 1990.
43Casali, Williges & Dryden, Effects of Recognition Accuracy and Vocabulary Size of a Speech Recognition System on Task Performance and User Acceptance, The Human Factors, pp. 183-196, Apr. 1990.
44 *Chaparro & Shufelt, Rational models for Quasi-Periodic Signals, IEEE Proceedings, vol. 74, No. 4, pp. 611 617, Apr. 1986.
45Chaparro & Shufelt, Rational models for Quasi-Periodic Signals, IEEE Proceedings, vol. 74, No. 4, pp. 611-617, Apr. 1986.
46Chasaide & Gobl, Contextual Variation of the Vowel Voice Source as a Function of Adjacent Consonants, Language and Speech, pp. 303-323.
47Chazan, Medan & Shvadron, Noise Cancellation for Hearing Aids, IEEE Transactions on Acoustics, Speech and Signal Processing, vol. 36, No. 11, pp. 1697-1705.
48 *Chen & Pan, Fast Search Algorithm for VQ-Based Recognition of Isolated Words, IEE Proceedings, Vol. 136, Pt 1, No. 6, pp. 391 396, Dec. 1989.
49Chen & Pan, Fast Search Algorithm for VQ-Based Recognition of Isolated Words, IEE Proceedings, Vol. 136, Pt 1, No. 6, pp. 391-396, Dec. 1989.
50 *Chen & Soong, Discriminative Training of High performance Speech Recognizer Using N Best Candidates, IEEE 1994.
51 *Chen, Soong & Lee, Large Vocabulary Word Recognition Based on Tree-Trellis Search, IEEE 1994.
52 *Cheng & O Shaughnessy, Short-Term Temporal Decomposition and its Properties for Speech Compression, IEEE Transactions on Signal Processing, vol. 39, No. 6, pp. 1282 1290, Jun. 1991.
53Cheng & O'Shaughnessy, Short-Term Temporal Decomposition and its Properties for Speech Compression, IEEE Transactions on Signal Processing, vol. 39, No. 6, pp. 1282-1290, Jun. 1991.
54Chips Recognize Speech, Electonics World & Wireless World, p. 137, Feb. 1990.
55Choi, Bang & Ann, A Robust Sequential Parameter Estimation for Time-Varying Speech Signal Analysis, IEEE 1994.
56 *Chou, Optimal Partitioning for Classification and Regression Trees, IEEE Transactions on Pattern Analysis & Machine Intelligence, vol. 13, No. 4, pp. 340 355, Apr. 1991.
57Chou, Optimal Partitioning for Classification and Regression Trees, IEEE Transactions on Pattern Analysis & Machine Intelligence, vol. 13, No. 4, pp. 340-355, Apr. 1991.
58Clery, Scottish Software May run Voice-Controlled Computer, Technology, Mar. 1990.
59Connolly, Edmonds, Guzy Johnson & Woodcock, Automatic Speech Recognition Based on Spectrogram Reading, International Journal of Man-Machine Studies, vol. 24, No. 6, pp. 611-621, Jun. 1986.
60Costlow, Board Heeds 1000 Words with 99.3% Accuracy, Electronic Design, p. 196, Oct. 16, 1986.
61Crochiere & Flannagen, Speech Processing: An Evolving Technology, AT&T Technica Journal, vol. 65, No. 5, pp. 2-11, Sep./Oct. 1986.
62Crystal & House, Articulation Rate and the Duration of Syllables and Stress Groups in Connected Speech, J. Acoustical Society of America, vol. 88, No. 1, pp. 101-112, Jul. 1990. [J. Acoustical Society of America, vol. 90, No. 2, p. 1191, Aug. 1991.
63Culling & Darwin, Perceptual Separation of Simultaneous Vowels: Within and Across-Formant Groupin by F0, J. Acoustical Society of America, vol. 93, No. 6, pp. 3454-3467, Jun. 1993.
64Damper, Voice-Input Aids for the Physically Disabled, International Journal of Man-Machine Studies, vol. 21, pp. 541-553, 1984.
65Das, Nadas, Nahamoo & Picheny, Adaptation Techniques for Ambience and Microphone Compensation in the IBM Tangora Speech Recognition System, IEEE 1994.
66 *De Mori, Lam & Gilloux, Learning and Plan Refinement in a Knowledge-Based System for Automatic Speech Recognition, IEEE Transactions on Pattern Analysis & Machine Intelligence, vol. PAMI-9, No. 2 pp. 289 305, Mar. 1987.
67De Mori, Lam & Gilloux, Learning and Plan Refinement in a Knowledge-Based System for Automatic Speech Recognition, IEEE Transactions on Pattern Analysis & Machine Intelligence, vol. PAMI-9, No. 2 pp. 289-305, Mar. 1987.
68De Mori, Larm & Probst, Rule-Based Detections of Speech Features for Automatic Speech Features for Automatic Speech Recognition Fundamentals in Computer Understanding, pp. 155-179.
69DeMori, Palakal & Cosi, Perceptual Models for Automatic Speech Recognition Systems, Advances in Computers, vol. 31, pp. 99-173.
70Derman, Recognizing Voices, Electronic Engineering Times, p. 39, Jan. 31, 1994.
71Dictating Greater Efficiency, The Engineer, pp 46, 49, Apr. 27, 1989.
72Difficult Speech-Recognition Technology Shows Signs of Maturity, Computer Design, pp. 23-29, Aug. 1, 1986.
73Dix & Bloodthooft, A New technique for Automatic Segmentation of Continuous Speech, NATO ASI Series, vol. F75, pp. 543-548, 1992.
74 *Doddington, Speaker Recognition-Identifying People by their Voices, IEE Proceedings, vol. 73, No. 11, pp. 1651 1663, Nov. 1985.
75Doddington, Speaker Recognition-Identifying People by their Voices, IEE Proceedings, vol. 73, No. 11, pp. 1651-1663, Nov. 1985.
76D'Orta, Ferretti, Martelli, Melecrinis, Scarci & Volpi, Large-Vocabulary Speech Recognition: A System for the Italian Language, IBM J. Res Develop, vol. 32, No. 2, pp. 217-228, Mar. 1988.
77 *Drews, Laroia, Pandel, Schumacher & St o lzle, CMOS Processor for Template-Based Speech-Recognition System, IEE Proceedings, vol. 136, Pt. 1, No. 2, pp. 155 161, Apr. 1989.
78Drews, Laroia, Pandel, Schumacher & Stolzle, CMOS Processor for Template-Based Speech-Recognition System, IEE Proceedings, vol. 136, Pt. 1, No. 2, pp. 155-161, Apr. 1989.
79DSP Boards Help Tackle a Tough Class of AL Tasks, Electronics, pp. 64-66, Aug. 21, 1986.
80 *Dumouchel, Suprasgmental Features and Continuous Speech Recognition, IEEE 1994.
81Dutoit, High Quality Test-To-Speech Synthesis: A Comparison of Four Candidate Algorithms, IEEE 1994.
82 *Eatock & Mason, A Quantitative Assessment of the Relative Speaker Discriminating Properties of Phonemes, IEEE 1994.
83Elman & Zipser, Learning the Hidden Structure of Speech, J. Acoustical Society of America, vol. 83, No. 4, pp. 1615-1626, Apr. 1988.
84 *Elman, A Personal Computer-based Speech Analysis and Synthesis System, IEEE MICRO, pp. 4 21, June 1987.
85Elman, A Personal Computer-based Speech Analysis and Synthesis System, IEEE MICRO, pp. 4-21, June 1987.
86 *Erell, Orgad & Goldstein, JND s in the LPC Poles of Speech and Their Application to Quantization of the LPC Filter, IEEE Transactions on Signal Processing, vol. 39, No. 2, pp. 308 318, Feb. 1991.
87Erell, Orgad & Goldstein, JND's in the LPC Poles of Speech and Their Application to Quantization of the LPC Filter, IEEE Transactions on Signal Processing, vol. 39, No. 2, pp. 308-318, Feb. 1991.
88Erkelens & Broersen, Analysis of Spectral Interpolation with Weighing Dependent on Fram Energy, IEEE 1994.
89 *Euler & Zinke, The Influence of Speech Coding Algorithms on Automatic Speech, IEEE 1994.
90Flores & Young, Continuous Speech Recognition in Noise Using Spectral Subtraction and HMM Adaptation, IEEE 1994.
91Fourakis, Tempo, Stress, and Vowel Reduction in American English, J. Acoustical Society of America, vol. 90, No. 4, pp. 1816-1827, Oct. 1991.
92Frankish, Decline in Accuracy of Automatic Speech Recognition as a Function of Time on Task: Fatigue or Voice Drift, International Journal of Man-Machine Studies, vol. 36, No. 6. pp. 797-816, Jun. 1992.
93Galanes, Savoji & Pardo, New Algorithm for Spectral Smoothing and Envelope Modification for LP-PSOLA Synthesis, IEEE 1994.
94 *Gao & Haton, A Hierarchical LPNN Network for Noise Reduction and Noise Degraded Speech Recognition, IEEE 1994.
95 *Garofolo, Robinson & Fiscus, The Development of File Formats for Very Large Speech Corpora: Sphere and Shorten, IEEE 1994.
96 *Georgoudis & Lagoyannis, Short-Time Spectrum Analyzer Based on Delta-Sigma Modulation, IEE Proceedings, vol. 133, Pt. G, No. 6, pp. 295 299, Dec. 1986.
97Georgoudis & Lagoyannis, Short-Time Spectrum Analyzer Based on Delta-Sigma Modulation, IEE Proceedings, vol. 133, Pt. G, No. 6, pp. 295-299, Dec. 1986.
98Glindemann & Dainty, Object Fitting to the Bispectral Phase by Using Least Squares, J. Opt. Soc. Am. A, vol. 10, No. 5, pp. 1056-1063, May 1993.
99 *Glinski, Lalimia, Cassiday, Koh, Gerveshi, Wilson & Kumar, The Graph Search Machine (GSM): A VLSI Architecture for Connected Speech Recognition and Other Applications, IEEE Proceedings, vol. 75, No 9, pp. 1170 1184, Sep. 1987.
100Glinski, Lalimia, Cassiday, Koh, Gerveshi, Wilson & Kumar, The Graph Search Machine (GSM): A VLSI Architecture for Connected Speech Recognition and Other Applications, IEEE Proceedings, vol. 75, No 9, pp. 1170-1184, Sep. 1987.
101Glinski, On the Use of Vector Quantization for Connected-Digit Recognition, AT&T Technical Journal, vol. 64, No. 5, pp. 1033-1045, May/Jun. 1985.
102 *Gong & Haton, Signal-to-String Conversation Based on High Likelihood Regions Using Embedded Dynamic Programming, IEEE Transactions on Pattern Analysis & Machine Intelligence, vol. 13, No. 3, pp. 297 302, Mar. 1991.
103Gong & Haton, Signal-to-String Conversation Based on High Likelihood Regions Using Embedded Dynamic Programming, IEEE Transactions on Pattern Analysis & Machine Intelligence, vol. 13, No. 3, pp. 297-302, Mar. 1991.
104 *Gong & Haton, Stochastic Trajectory Modeling for Speech Recognition, IEEE 1994.
105Grant, Ardell, Kuhl & Sparks, The Contribution of Fundamental Frequency, Amplitude Envelope, and Voicing Duration Cues to Speechreading in Normal-Hearing Subjects, J. Acoustical Society of America vol. 77, No. 2, pp. 671-677, Feb. 1985.
106Grossberg & Wyse, A Neural Network Architecture for Figure-Ground Separation of Connected Scenic Figures, Neural Networks, vol. 4, pp. 723-742, 1991.
107Gupta, Lennig & Mermeistein, Fast Search Strategy in a Large Vocabulary Word Recognizer, J. Acoustical Society of America, vol. 84, No. 6 pp. 2007-2017, Dec. 1988.
108Handheld Computer Follows Voice Commands, Design News, p. 38, Mar. 7, 1988.
109Harris, On The Use of Windows for Harmonic Analysis with the Discrete Fourier Transform, Proceeding of the IEEE, vol. 66, No. 1, pp. 51-83, Jan. 1978.
110Hasegawa & Hata, Fundamental Frequency as an Acoustic Cue to Accent Perception, Language and Speech, vol. 35, No, 1,2, pp. 87-98, 1992.
111 *Hayakawa & Itakura, Text-Dependent Speaker Recognition Using the Information in the Higher Frequency Band, IEEE 1994.
112 *Hemando & Nadeu, Speech Recognition in Noisy Car Environment Based on OSALPC Representation and Robust Similarity Measuring Techniques, IEEE 1994.
113Hermes & vanGestel, The Frequency Scale of Speech Intonation, J. Acoustical Society of America, vol. 90, No. 1, pp. 97-102, Jul. 1991.
114Holmgren, Toward Bell System Applications of Automatic Speech Recognition, The Bell System Technical Journal, vol. 62, No. 6, pp. 1865-1879, Jul./Aug. 1983.
115Howell & Williams, Acoustic Analysis and Perception of Vowels in Children's and Teenager's Stuttered Speech, J. Acoustical Society of America, vol. 91, No. 3, pp. 1697-1706, Mar. 1992.
116 *Hunt, A Generalized Model for Utilizing Prosopic Information in Continuous Speech Recognition, IEEE 1994.
117 *Hurst & Brodersen, An MOS-LSI Autocorrelator for Linear Prediction of Speech, IEEE Journal of Solid-State Circuits, vol. sc-19, No. 6, pp. 1022 1029, Dec. 1984.
118Hurst & Brodersen, An MOS-LSI Autocorrelator for Linear Prediction of Speech, IEEE Journal of Solid-State Circuits, vol. sc-19, No. 6, pp. 1022-1029, Dec. 1984.
119I Recognize that Voice!, Electronics & Power, p. 783, Nov/Dec. 1986.
120Immerseel & Martens, Pitch and Voiced/Unvoiced Determination with an Auditory Model, J. Acoustical Society of America, vol. 91, No. 6, pp. 3511-3526, Jun. 1992.
121Ince, Digital Speech Processing, 1992.
122 *Jack, Laver & Blauert, Editorial: Speech Technology, IEE Proceedings, vol. 136, Pt. 1, No. 2, p. 109, Apr. 1989.
123Jakatdar & Mulla, Speech Communication for Personal Computers, Electrical Communication, vol. 60, No. 1., pp. 79-86, 1986.
124 *James & Young, A Fast Lattice-Based Approach to Vocabulary Independent Wordspotting, IEEE 1994.
125 *Jeanrenaud, Siu, Rohlicek, Meteer & Gish, Spotting Events in Continuous Speech, IEEE 1994.
126 *Jelinek, The Development of an Experimental Discrete Dictation Recognizer, IEE Proceedings, vol. 73, No. 11,pp. 1616 1624, Nov. 1985.
127Jelinek, The Development of an Experimental Discrete Dictation Recognizer, IEE Proceedings, vol. 73, No. 11,pp. 1616-1624, Nov. 1985.
128Jensen, High Frequency Phase Response Specifications--Useful or Misleading?, J. Ado Eng. Soc., vol. 36, No. 12, pp. 968-975, Dec. 1988.
129 *Johansen & Johnsen, Non-Linear Input transformations for Discriminative HMMS, IEEE 1994.
130John Gallant, Voice-Recognition System Learns User's vocabularies and Speaking Styles, EDN p. 106, Ma 24,1990.
131Josenhans, Lynch, Rogers, Rosinski & VanDame, Speech Processing Application Standards, AT&T Technica Journal, vol. 65, No. 5, pp. 23-33, Sep./Oct. 1986.
132Jouny & Moses, The Bispectrum of Complex Signals Definitions and Properties, IEEE, pp. 2833-2837, IEEE Transactions on Signal Processing, vol. 40, No. 11, pp. 2833-2836, Nov. 1992.
133Juang, Maximum-Likelihood Estimation for Mixture Multivariate Stochastic Observations of Markov Chains, AT&T Technical Journal, vol. 64, No. 6, pp. 1235-1249, Aug. 1985.
134Junqua, The Lombard Reflex and its Role on Human Listeners and Automatic Speech Recognizers, J. Acoustical Society of America, vol. 91, No. 1, pp. 510-524, Jan. 1993.
135 *Kao, Hemphill, Wheatley & Rajasekaran, Toward Vocabulary Independent Telephone Speech Recognition ,IEEE 1994.
136 *Kataoka, Moriya & Hayahi, Implementation and Performance of an 9-kbit/s Conjugate Structure Celp Speech Coder, IEEE 1994.
137Kauderer, Becker & Powers, Acousto-optical Bispectral Processing, Applied Optics, vol. 28, No. 3, pp. 627=14 637, Feb. 1, 1989.
138Kenny, Labute, Li, & O'Shaughnessy, New Graph Search Techniques for Speech Recognition, IEEE 1994.
139Keurs, Festen & Plomp, Effect of Spectral Envelope Smearing on Speech Receptions, II, J. Acoust Soc. Am. vol. 93, No. 3, pp. 1547-1552, Mar. 1993.
140Kikuta, Iwata & Nagata, Distance Measurement by the Wavelength Shift of Laser Diode Light, Applied Optics, vol. 25, No. 17, pp. 2976-2980, Sep. 1, 1988.
141Kim & Hayes, Phase Retrieval Using a Window Function, IEEE Transactions on Signal Processing, vol. 41, No. 3, pp. 1409=14 1415, Mar. 1993.
142Kim & Hayes, Phase Retrieval Using Two Fourier-Transform Intensities, J. Opt. Soc. Am. A, vol. 7, No. 3, pp. 441-449, Mar. 1990.
143Kluender, Effects of First Formant Onset Properties on Voicing Judgments Result From Processes Not Specific to Humans, J. Acoustical Society of America, vol. 90, No. 1, pp. 83-96, Jul. 1991.
144Kniffen, Becher & Powers, Bispectral magnitude and Phase Recovery Using a Wide Bandwidth Acousto- Optic Processor, Applied Optics, vol. 3, No. 8, pp. 1015-1029, Mar. 10, 1992.
145Kobatake & Matsunoo, Degraded Word Recognition Based on Segmental Signal-To-Noise Ratio Weighing, IEEE 1994.
146 *Kobayashi, Mine & Shirai, Markov Model Based Nowise Modeling and Its Application to Noisy Speech Recognition using Dynamical Features of Speech, IEEE 1994.
147Koenig, Spectrographic Voice Identification: A Forensic Survey, J. Acoustical Society of America, vol. 79, No. 6, pp. 2088-2090, Jun. 1986.
148Kohata & Takagi, Vector Quantization With Hyper-Columnar Clusters, IEEE 1994.
149 *Kompe, Batliner, Keissling, Kilian, Niemann & N o th, Automatic Classification of Prosopically Marked Phrase Boundaries in German, IEEE 1994.
150Kompe, Batliner, Keissling, Kilian, Niemann & Noth, Automatic Classification of Prosopically Marked Phrase Boundaries in German, IEEE 1994.
151 *Kong & Kosko, Differential Competitive Learning for Centroid Estimation and Phoneme Recognition, IEEE Transactions on Neural Networks, vol. 2, No. 1, pp. 118 124, Jan. 1991.
152Kong & Kosko, Differential Competitive Learning for Centroid Estimation and Phoneme Recognition, IEEE Transactions on Neural Networks, vol. 2, No. 1, pp. 118-124, Jan. 1991.
153 *Kosaka & Sagayama, Tree-Structured Speaker Clustering for Fast Speaker Adaptation, IEEE 1994.
154 *Krishnan & Rao, Segmental Phoneme Recognition Using Piecewise Linear Regression, IEEE 1994.
155 *Krubsack & Niederjohn, An Autocorrelation Pitch Detector and Voicing Decision with Confidence Measures Developed for Noise-Corrupted Speech, IEEE Transactions on Signal Processing, vol. 39, No. 2 pp. 319 329, Feb. 1991.
156Krubsack & Niederjohn, An Autocorrelation Pitch Detector and Voicing Decision with Confidence Measures Developed for Noise-Corrupted Speech, IEEE Transactions on Signal Processing, vol. 39, No. 2 pp. 319-329, Feb. 1991.
157Kuabara, An Approach to Normalization of Coarticulation Effects for Vowels in Connected Speech, J. Accoustical Society of America, pp. 686-694, Feb. 1985.
158Kubala, Anastasakos, Makhoul, Nguyen, Schwartz & Zavaliagkos, Comparative Experiments on Large Vocabulary Speech Recognition, IEEE 1994.
159Kurtzberg, Feature Analysis for Symbol Recognition by Elastic Matching, IBM J. Res Develop, vol. 31, No. 1, pp. 91-95, Mar. 1987.
160Kurzweil, Better Speech Recognition Means that Computers Must Mimic the Human Brain, Electronic Design, pp. 83-84, Nov. 15, 1984.
161Kurzweil, Beyond Pattern Recognition, BYTE, pp. 277-288, Dec. 1989.
162Kurzweil, The Technology of the Kurzweil Voice Writer, BYTE, pp. 177-186, Mar. 1986.
163Laface & DeMori, Speech Recognition and Understanding, Series F. Computer and Systems Sciences vol. 75, 1991.
164Langhans & Kohlrausch, Differences in Auditory Performance Between Monaural and Diotic Conditions, I: Masked thresholds in Frozen Noise, J. Acoustical Society of America, vol. 91, No. 6, pp. 3456-3470, Jun. 1992.
165Laprice & Berger, A New Paradigm for Reliable Automatic Formant Tracking, IEEE 1994.
166Leary & Morgan, Fast and Accurate Analysis with LPC Gives a DSP Chip Speech-Processing Power, Electronic Design, pp. 153-158, Apr. 17, 1986.
167Lee, Automatic Speech Recognition, 1989.
168Lee, Hauptmann & Rudnicky, The Spoken Word, BYTE, pp. 225-232, Jul. 1990.
169 *Lee, Information-Theoretic Distortion measures for Speech Recognition, IEEE Transactions on Signal Processing, Vol. 39, No. 2, pp. 330 335, Feb. 1991.
170Lee, Information-Theoretic Distortion measures for Speech Recognition, IEEE Transactions on Signal Processing, Vol. 39, No. 2, pp. 330-335, Feb. 1991.
171Leinonen, Hiltunen, Torkkola, & Kangas, Self-Organized Acoustic Feature Map in Detection of Fricative-Vowel Coarticulation, J. Acoustical Society of America, vol. 93, No. 6, pp. 3468-3474, Jun. 1993.
172Lennig & Mermeistein, Use of Vowel Duration Information in a Large Vocabulary Word Recognizer, J. Acoustical Society of America, vol. 86, No. 2, pp. 540-548, Aug. 1989.
173 *Levinson & Roe, A Perspective on Speech Recognition IEEE Communications Magazine, pp. 28 34, Jan. 1990.
174Levinson & Roe, A Perspective on Speech Recognition--IEEE Communications Magazine, pp. 28-34, Jan. 1990.
175Levinson, Rabiner & Sondhi, An Introduction to the Application of the Theory of Probabilistic Functions of a Markov Process to Automatic Speech Recognition, The Bell System Technical Journal, vol. 62, No. 4 pp. 1035-1074, Apr. 1983.
176Levinson, Recognition of Continuous Complex Speech by Machine, J. Acoustical Society of America, vol. 87, No. 1, pp. 422-423, Jan. 1990.
177 *Levinson, Structural Methods in Automatic Speech Recognition, IEE Proceedings, vol. 73, No. 11, pp. 1625 1650, Nov. 1985.
178Levinson, Structural Methods in Automatic Speech Recognition, IEE Proceedings, vol. 73, No. 11, pp. 1625-1650, Nov. 1985.
179Lindholm, Dorman, Taylor & Hannley, Stimulus Factors Influencing the Identification of Voiced Stop Consonants by Normal-Hearing and Hearing-Impaired Adults, J. Acoustical Society of America,vol. 83 No. 4, pp. 1608-1614, Apr. 1988.
180Linear Prediction of Speech.
181 *Lippmann, Pattern Classification Using Neural Networks, IEEE Communications Magazine, pp. 47 64, Nov. 1989.
182Lippmann, Pattern Classification Using Neural Networks, IEEE Communications Magazine, pp. 47-64, Nov. 1989.
183 *Liu, Lee, Wang & Chang, Layered Neutral Nets Applied in the Recognition of Voiceless Unaspirated Stops, IEE Proceedings, vol. 136, Pt. 1,No. 2, pp. 69 75, Apr. 1989.
184Liu, Lee, Wang & Chang, Layered Neutral Nets Applied in the Recognition of Voiceless Unaspirated Stops, IEE Proceedings, vol. 136, Pt. 1,No. 2, pp. 69-75, Apr. 1989.
185 *Liu, Stern, Acero & Moreno, Environment normalization for Robust Speech Recognition using Direct Cepstral Comparison, IEEE 1994.
186Lockwood & Alexandre, Root Adaptive Homomorphic Deconvolution Schemes For Speech Recognition in Noise, IEEE 1994.
187Lofqvist, Acoustic and Aerodynamic Effects of Interarticulator Timing in Voiceless Consonants, Language and Speech, vol. 35, No, 1,2, pp. 15-28, 1992.
188 *Lowe & Webb, Optimized Feature Extraction and the Bayes Decision in Feed-Forward Classifier Networks, IEEE Transactions on Pattern Analysis & Machine Intelligence, vol. 13, No. 4, pp. 355 364, Apr. 1991.
189Lowe & Webb, Optimized Feature Extraction and the Bayes Decision in Feed-Forward Classifier Networks, IEEE Transactions on Pattern Analysis & Machine Intelligence, vol. 13, No. 4, pp. 355-364, Apr. 1991.
190Lyons, Chirp-Z Transform Efficiently Computes Frequency Spectra, EDN, pp. 161-170, May 25, 1989.
191Madhavan, Minimal Repetition Evoked Potentials by Modified Adaptive Line Enhancement, IEEE Transactions on Biomedical Engineering, vol. 39, No. 7, pp. 760-764, Jul. 1992.
192Mandel, A Commercial Large-Vocabulary Discrete Speech Recognition System: Dragon Dictate, Language and Speech, vol. 35, No, 1,2, pp. 237-246, 1992.
193Marchal & Hardcastle, Accor. Instrumentation and Database for the Cross-Language Study of Coarticulation, Language and Speech, vol. 36, No. 2, 3, pp. 137-153, 1993.
194 *Mariani, Hamlet A Prototype of a Voice-Activated Typewriter, IEE Proceedings, vol. 136, Pt. 1, No. 2, pp.162 166, Apr. 1989.
195Mariani, Hamlet A Prototype of a Voice-Activated Typewriter, IEE Proceedings, vol. 136, Pt. 1, No. 2, pp.162-166, Apr. 1989.
196Marron Sanchez & Sullivan, Unwrapping Algorithm for Least-Squares Phase Recovery from the Modula 2πBispectrum Phase, J. Opt. Sec. Am. A, vol. 7, No. 1 , pp. 14-20, Jan. 1990.
197 *Martinelli, Orlandi, Ricotti & Ragazzini, Identification of Stable Nonstationary Lattice Predictors by Linear Programming, IEE Proceedings, vol. 74, No. 5, pp. 759 776, May 1986.
198Martinelli, Orlandi, Ricotti & Ragazzini, Identification of Stable Nonstationary Lattice Predictors by Linear Programming, IEE Proceedings, vol. 74, No. 5, pp. 759-776, May 1986.
199 *Martino, Mari, Mathieu, Perot & Smaili, Which Model for Future Speech Recognition Systems: Hidden Markov Models or Finite-State Automata IEEE 1994.
200Martino, Mari, Mathieu, Perot & Smaili, Which Model for Future Speech Recognition Systems: Hidden Markov Models or Finite-State Automata? IEEE 1994.
201Matsuoka & Ulrych, Phase Estimation Using the Bispectrum, Proceedings of the IEEE, vol. 72, No. 10, pp. 1403-1411, Oct. 1984.
202McGrath & Summerfield, Internodal Timing Relations and Audio-Visual Speech Recognition by Normal-Hearing Adults, J. Acoustical Society of America, vol. 77, No. 2, pp. 678-685, Feb. 1985.
203McKee, TMS32010 Routine Finds Phase, EDN, p. 148, May 10, 1990.
204 *Mclnnes, Jack & Laver Template Adaptation in an Isolated Word-Recognition System, IEE Proceedings, vol. 136, Pt. 1, No. 2, pp. 119 126 Apr. 1989.
205Mclnnes, Jack & Laver Template Adaptation in an Isolated Word-Recognition System, IEE Proceedings, vol. 136, Pt. 1, No. 2, pp. 119-126 Apr. 1989.
206Meisel, Talk to your Computer, BYTE, pp. 113-120, Oct. 1993.
207 *Meng, Seneff & Zue, Phonological Parsing for Reversible Letter-to-Sound/Sound-to-Letter Generation, IEEE 1994.
208Mercer & Cohen, A Method for Efficient Storage and Rapid Application of Context-Sensitive Phonologica Rules for Automatic Speech recognition, IBM J. RES DEVELOP vol. 31, No. 1, pp. 81-90, Mar. 1987.
209Mercer, Statistical Modeling for Automatic Speech Recognition, AFIPS Conference Proceedings, May 16-19 1983, p. 643.
210 *Mercier, Bigorgne, Miclet, Le Guennec & Querre, Recognition of Speaker-Dependent Continuous Speech with KEAL, IEEE Transaction on Signal Processing, vol. 136, PT. 1, No. 2, pp. 145 154, Apr. 1989.
211Mercier, Bigorgne, Miclet, Le Guennec & Querre, Recognition of Speaker-Dependent Continuous Speech with KEAL, IEEE Transaction on Signal Processing, vol. 136, PT. 1, No. 2, pp. 145-154, Apr. 1989.
212Merialdo, Multilevel Decoding for Very-Large-Size-Dictionary Speech Recognition, IBM J. Res Develop, vol. 32, No. 2, pp. 227-237, Mar. 1988.
213Milenkovic, Voice Source Model for Continuous Control of Pitch Period, J. Acoustical Society of America, vol. 93, No. 2, pp. 1087-1096, Feb. 1993.
214 *Milner & Vaseghi, Speech Modeling Using Cepstral-Time Feature Matrices and Hidden Markov Models, IEEE 1994.
215 *Minami, Shikano, Takahashi & Yamada, Search Algorithm That Merges Candidates In Meaning Level Fo Very Large Vocabulary Spontaneous Speech Recognition, IEEE 1994.
216Mizuno & Abe, Voice Conversion Based on Piecewise Linear Conversion Rules of Formant Frequency and Spectrum Tilt, IEEE 1994.
217 *Mohan & Komandur, Performance of a Multiprocessor-Based Parallel Stack Algorithm Speech Encoder, IEEE, pp. 463 467, 1987.
218Mohan & Komandur, Performance of a Multiprocessor-Based Parallel Stack Algorithm Speech Encoder, IEEE, pp. 463-467, 1987.
219 *Mohan, Lin & Kryskowski, Stack Algorithm Speech Encoding with Fixed and Variable Symbol Release Rules, IEEE Transactions on Communications, vol. Com-33, No. 9, pp. 1015 1019, Sep. 1985.
220Mohan, Lin & Kryskowski, Stack Algorithm Speech Encoding with Fixed and Variable Symbol Release Rules, IEEE Transactions on Communications, vol. Com-33, No. 9, pp. 1015-1019, Sep. 1985.
221Moore, Peters & Glasberg, Detections of Temporal Gaps in Sinusoids: Effects of Frequency and Level, J. Acoust. Soc. Am. vol. 93, No. 3, pp. 1563-1570, Mar. 1993.
222 *Moreno & Stem, Sources of Degradation of Speech Recognition in the Telephone Network, IEEE 1994.
223Moss & Simmons, Acoustic Image Representation of a Point Target in the Bat Eptesicus fuscus: Evidence for Sensitivity to Echo Phase in Bat Sonar, J. Acoust. Soc. Am., vol. 93, No. 3, pp. 1553-1562, Mar. 1993.
224Mullick & Reddy, Channel Characterization Using Bispectral Analysis, Proceedings of the IEEE, vol. 76, No. 1, pp. 88-89, Jan. 1988.
225 *Murano, Unagami & Amano, Echo Cancellation and Applications, IEEE Communications Magazine, pp. 49 55, Jan. 1990.
226Murano, Unagami & Amano, Echo Cancellation and Applications, IEEE Communications Magazine, pp. 49-55, Jan. 1990.
227Naegele, Graphics Tablet Tales to Compete with Mouse, Electronics, p. 14, Oct. 1986.
228 *Naik, Speaker Verification: A Tutorial, IEEE Communications Magazine, pp. 42 48, Jan. 1990.
229Naik, Speaker Verification: A Tutorial, IEEE Communications Magazine, pp. 42-48, Jan. 1990.
230Nakajima, Phase Retrieval Using The Logarithmic Hilbert Transform and the Fourier-Series Expansion, J. Opt. Sec. Am. A, vol. 5, No. 2, pp. 257-262, Feb. 1988.
231Nakajima, Signal-to-Noise Ratio of the Bispectral Analysis of Speckle Interferometry, J. Opt. Am. A., vol. 5, No. 9, pp. 1477-1491, Sep. 1988.
232 *Nandi, Aburdene, Constantindes, & Dologlou, Variation of Vector Quantisation and Speech Waveform Coding, IEE Proceedings-1, vol. 138, No. 2, pp. 76 80, Apr. 1991.
233Nandi, Aburdene, Constantindes, & Dologlou, Variation of Vector Quantisation and Speech Waveform Coding, IEE Proceedings-1, vol. 138, No. 2, pp. 76-80, Apr. 1991.
234 *Nandkumar & Hansen, Speech Enhancement Based on a New Set of Auditory Constrained Parameters, IEEE 1994.
235Neumeyer & Weintraub, Probablistic Optimum Filtering for Robust Speech Recognition, IEEE 1994.
236Newell, Arnott, Dye & Caims, A Full-Speed Listening Typewriter Simulation, International Journal of Man- Machine Studies, vol. 35, No. 2, pp. 119-131, Aug. 1991.
237Newman, Detecting Speech with an Adaptive Neural Network, Electronic Design, pp. 79-90, Mar. 22, 1990.
238 *Ney, Dynamic Programming Parsing for Context-Free Grammars in Continuous Speech Recognition, IEEE Transactions on Signal Processing, vol. 39, No. 2, pp. 336 340, Feb. 1991.
239Ney, Dynamic Programming Parsing for Context-Free Grammars in Continuous Speech Recognition, IEEE Transactions on Signal Processing, vol. 39, No. 2, pp. 336-340, Feb. 1991.
240Nikais & Liu, Bicepstrum Computation Based on Second- and Third-Order Statistics with Applications, IEEE, pp. 2381-2385, 1990.
241Nikais, ARMA Bispectrum Approach to Nonminimum Phase System Identification, IEEE Transactions on Acoustics, and Signal Processing, vol. 36, No. 4, pp. 513-524, Apr. 1988.
242Nikias & Chiang, Higher-Order Spectrum Estimation Via Noncausal Autoregressive modeling and Deconvolution, IEEE. Transactions on Acoustics, Speech and Signal Processing, vol. 36, No. 12, pp. 1911-1913 Dec. 1988.
243Nikias & Pan, ARMA Modeling of Fourth-Order Cumulants and Phase Estimation, Circuits Systems Signal Process, vol. 7, No. 3, pp. 291-325, 1988.
244Nikias & Raghuveer, Bispectrum Estimation: A Digital Signal Processing Framework, Proceeding of the IEEE, vol. 75, No. 7, pp. 869-891, Jul. 1987.
245 *Niles, Acoustic Modeling for Speech Recognition Based on Spotting of Phonetic Units, IEEE 1994.
246Niranjan, Recursive Tracking of Formants in Speech Signals, IEEE 1994.
247Noda & Shirazi, A MRF-Based Parallel Processing Algorithm for Speech Recognition Using Linear Predictive HMM, IEEE 1994.
248Noonan, Premus & Irza, AR Model Order Selection Based on Bispectral Cross Corretion, IEEE Transactions on Signal Processing, Vol 39, No. 6, pp. 1440-1443, June 1991.
249Nossair & Zahorian, Dynamic Spectral Shape Features as Acoustic Correlates for Initial Stop Consonants, J. Acoustical Society of America, vol. 89, No. 6, pp. 2978-2991, Jun. 1991.
250O'Brien, Knowledge-Based Systems in Speech Recognition: A Survey of, International Journal of Man-Machin Studies, vol. 38, No. 1, pp. 71-95, Jan. 1993.
251O'Brien, Spectral Features of Plosives in Connected-Speech Signals, International Journal of Man-Machine Studies, vol. 38, pp. 97-127, 1993.
252Ohala, Coarticulation and Phonology, Language and Speech, vol. 36, No. 2, 3, pp. 155-170, 1993.
253 *Ohmura, Fine Pitch Contour Extraction by Voice Fundamental Wave Filtering Method, IEEE 1994.
254O'Kane & Kenne, Sidebar 1: Automatic Speech Recognition: One of the Hard Problems of Artificial Intelligence, Library Hi Tech, Issue 37-38, pp. 42-49, 1992.
255One-Card System Recognizes Words in a Sentence with 90% Accuracy, Electronics Week, pp. 17-19, Oct. 15, 1984.
256 *Openshaw & Mason, On The Limitations of Cepastral Features in Noise, IEEE 1994.
257Oppenheim, Digital Processing of Speech, Digital Processing in Audio Signals, pp. 117-168.
258Owens, Signal Processing of Speech, McGraw-Hill, Inc., 1993, 1-9, 35-39, 70-72, 74-80, 85-87.
259 *Pal & Mirta, Multilayer Perception, Fuzzy Sets, and Classification, IEEE Transactions on Neural Networks, vol. 3, No. 5, pp. 683 697, Sep. 1992.
260Pal & Mirta, Multilayer Perception, Fuzzy Sets, and Classification, IEEE Transactions on Neural Networks, vol. 3, No. 5, pp. 683-697, Sep. 1992.
261Pallett, Performance of Research of the National Bureau of Standards, Journal of Research of the National Bureau of Standard, vol. 90, No. 5, pp. 371-387, Oct. 1985.
262 *Pawate & Dowling, A New Method for Segmenting Continuous Speech, IEEE 1994.
263PC Recognizes 20,000 Spoken Words, Machine Design, p. 16, May 7, 1987.
264Peacocke & Graf, An Introduction to Speech and Speaker Recognition, Computer, pp. 26-33, Aug. 1990.
265Perdue & Rissanen, Conversant 1 Voice System: Architecture and Applications, AT&T Technical Journal, vol. 65, No. 5, pp. 34-47, Sep./Oct. 1986.
266Perez-Ilzarbe, Phase Retrieval From the Power Spectrum of a Periodic Object, J. Opt. Soc. Am. A., vol. 9, No. 12, pp. 2138-2148, Dec. 1992.
267Pezeshki, Elgar, Krishna & Burton, Auto and Cross-Bispectral Analysis of a System of Two Coupled Oscillators With Quadratic Nonlinearities Possessing Chaotic Motion, Journal of Applied Mechanics, vol. 59, pp. 657=14 663, Sep. 1992.
268Pfingst, DeHann & Holloway, Stimulus Features Affecting Psychophysical Detection Thresholds for Electrical Stimulation of the Cochlea. I: Phase Duration and Stimulus Duration, J. Acoustical Society o America, vol. 90, No. 4, pp. 1857-1866, Oct. 1991.
269Phillip, Applications of Automatic Speech Recognition and Synthesis in Libraries and Information Services: A Future Scenario, Library Hi Tech, Issue 35, pp. 89-93, 1991.
270Pisoni, Nusbaum & Greene, Perception of Synthetic Speech Generated by Rule, IEE Proceedings, vol. 73, No. 11, p. 1665, Nov. 1985.
271 *Pisoni, Nusbaum & Greene, Perception of Synthetic Speech Generated by Rule, IEE Proceedings, Vol. 73, No. 11, pp. 1665-, Nov. 1985.
272Pullin, Developing Systems to Hear Through the Shopfloor Din, The Engineer, pp. 32-33, Sep. 21, 1989.
273Qi & Fox, Analysis of Nasal Consonants Using Perceptual Linear Prediction, J. Accoustical Society of America, pp. 1718-1726, Mar. 1992.
274Qi & Shipp, An Adaptive Method for Tracking Voicing Irregularities, J. Acoustical Society of America, vol. 91, No. 6, pp. 3471-3475, Jun. 1992.
275 *Quenot, Gauvain, Gangolf & Mariani, A Dynamic Programming Processor for Speech Recognition, IEEE Journal of Solid-State Circuits, vol. 24, No. 2, pp. 349 357, Apr. 1989.
276Quenot, Gauvain, Gangolf & Mariani, A Dynamic Programming Processor for Speech Recognition, IEEE Journal of Solid-State Circuits, vol. 24, No. 2, pp. 349-357, Apr. 1989.
277Rabiner, Juang, Levinson & Sondhi, Some Properties of Continuous Hidden Markov Model Representations, AT&T Technical Journal, vol. 64, No. 6, pp. 1251-1270, Aug. 1985.
278Rabiner, Levinson, & Sondhi, On the Application of Vector Quantization and Hidden Markov Models to Speaker-Independent, isolated Word Recognition, The Bell System Technical Journal, vol. 62, No. 4, pp 1075-1105, Apr. 1983.
279Rabiner, On the Application of Energy Contours to the Recognition of Connected Word Sequences, AT& Technical Journal, vol. 63, No. 9, pp. 1981-1995, Nov. 1985.
280Rabiner, Schafer, Digital Processing of Speech Signals, Prentice-Hall Inc., 1978, pp. 38-47, 116-123, 130-135, 462-463, 489-490.
281Rabiner, Wilpon, & Juang, A Sgmental κ-Means Training Procedure for Connected Word Recognition, AT&T Technical Journal, vol. 65, No. 3, pp. 21-31, May/Jun. 1986.
282Rahim & Juang, Signal Bias Removal for Robust Telephone Based Speech Recognition in Adverse Environments, IEEE 1994.
283Rao & Gabr, An Introduction to Bispectral Analysis and Bilinear Time Series Models, Book Reviews, pp. 326=14 329.
284 *Rashwan & Fahmy, New Technique for Speaker-Independent Isolated-Work Recognition, IEEE Proceedings, vol. 135, Pt. F, No. 3, pp. 251 546, Jun. 1988.
285Rashwan & Fahmy, New Technique for Speaker-Independent Isolated-Work Recognition, IEEE Proceedings, vol. 135, Pt. F, No. 3, pp. 251-546, Jun. 1988.
286Rayfield & Silverman, An Approach to DFT Calculations Using Standard Microprocessors, IBM J. Res. Develop, vol. 29, No. 2, pp. 170-176, Mar. 1985.
287Repp, Perception of the [m]-[n] Distinction in CV Syllables, J. Accoustical Society of America, pp. 1987-1999 Jun., 1986.
288Revoile, Kozma-Spytek, Holden-Pitt, Pickett & Droge, VCVs vs CVCs for Stop/Fricative Distinctions by Hearing-Impaired and Normal-Hearing Listeners, J. Acoustical Society of America, vol. 89, No. 1, pp. 457-406, Jan. 1991.
289Rosenberg, Speech Processing: Hearing Better, Talking More, Electronics, pp. 26-30, Apr. 21, 1986.
290Saffari, Putting DSPs to Work BYTE, pp. 259-272, Dec. 1989.
291 *Sagisaka, Speech Synthesis from Text, IEEE Communications Magazine, pp. 35 41, Jan. 1990.
292Sagisaka, Speech Synthesis from Text, IEEE Communications Magazine, pp. 35-41, Jan. 1990.
293Saito, Speech Science and Technology, IOS Press.
294 *Samouelian, Knowledge Based Approach to Consonant Recognition, IEEE 1994.
295Schmidbauer, Casacuberta, Castro, Hegerl, Hoge, Sanchez & Zlokamik, Articulatory Representation and Speech Technology, Language & Speech-1993, 36 (2,3) 331-351.
296Schmidbauer, Casacuberta, Castro, Hegerl, Hoge, Sanchez & Zlokarnik, Articulatory Representation and Speech Technology, Language and Speech, vol. 36, No, 2,3, pp. 331-351, 1993.
297Schmidt-Neilsen & Stern, Identification of Known Voices as a Function of Familiarity and Narrow-Band Coding, J. Acoustical Society of America, vol. 77, No. 2, pp. 658-670, Feb. 1985.
298 *Schroeder, Linear Predictive Coding of Speech: Review and Current Directions, IEEE Communications Magazine, pp. 54 61, Aug. 1985.
299Schroeder, Linear Predictive Coding of Speech: Review and Current Directions, IEEE Communications Magazine, pp. 54-61, Aug. 1985.
300Scoring 98.6% in Speech Recognition, Electronics, p. 41, Oct. 2, 1986.
301 *Seide & Mertins, Non-Linear Regression Based Feature Extraction For Connected-Work Recognition in Noise, IEEE 1994.
302Shigeno, Assimilation and Contrast in the Phonetic Perception of Vowels, J. Accoustical Society of America, pp. 103-111, Jul. 1991.
303 *Shimodaira & Nakai, Prosopic Phrase Segmentation By Pitch Pattern Clustering, IEEE 1994.
304Shriberg, Perceptual Restoration of Filtered Vowels with Added Noise, Language and Speech, vol. 35, No, 1,2, pp. 127-136, 1992.
305Simpson, McCauley, Roland, Ruth & Williges, System Design for Speech Recognition and Generation, The Human Factors, Apr. 1990.
306 *Slaney, Naar & Lyon, Auditory Model Inversion for Sound Separation, IEEE 1994.
307Smart Cards will Respond to Owner's Voice, Radio-Electronics.
308Smarte & Penney, Sounds and Images, BYTE, pp. 243-248., Dec. 1989.
309Sole, Phonetic and Phonological Processes: The Case of Nasalization, Language and Speech, 1992, 35 (1,2), 29-43.
310Soltis, Automatic Identification Systems: Strengths, Weaknesses and Future Trends, IE, pp. 55-59, Nov. 1985.
311Sommers, Moody, Prosen & Stebbins, Formant Frequency Discrimination by Japanese Macaques (Macac Fuscata) J. Acoustical Society of America, vol. 91, No. 6, pp. 3499-3509, Jun. 1992.
312Speech I/O Products Offer Board-level Solutions, Computer Design, pp. 36-40, Mar. 15, 1986.
313Speech Recognition Problems Examined, Society of Automotive Engineers, vol. 95, No. 8, pp. 59-61, 1987.
314Speech Recognition Trial, Monitor, vol. 28, No. 5, p. 227, Jun. 1986.
315Speech-Recognition Products, EDN, pp. 112-122, Jan. 19, 1989.
316Springer, Sliding FFT Computers Frequency Spectra in Real Time, EDN, pp. 161-170, Sep. 29, 1988.
317Stommen, Talking Back to Big Bird: Preschool users and a Simple Speech Recogniition System, ETR&D, vol. 41, No. 1, pp. 5-16.
318 *Suaudeau, An Efficient Combination of Acoustic and Supra-Segmental Information in a Speech Recognition System, IEEE 1994.
319Sundberg Lindblom & Liljencrants, Formant Frequency Estimates for Abruptly Changing Area Functions A Comparison Between Calculations and Measurements, J. Acoustical Society of America, vol. 91, No. 6 pp. 3478-3482, Jun. 1992.
320 *Sutherland, Jack & Laver, Improved Pitch Detection Algorithm Employing Temporal Structure Investigation of the Speech Waveform, IEEE Proceedings, vol. 135, Pt. F, No. 2, pp. 169 174, Apr. 1988.
321Sutherland, Jack & Laver, Improved Pitch Detection Algorithm Employing Temporal Structure Investigation of the Speech Waveform, IEEE Proceedings, vol. 135, Pt. F, No. 2, pp. 169-174, Apr. 1988.
322Takebayashi & Kanazawa, Adaptive Noise Immunity Learning for Word Spotting, IEEE 1994.
323 *Tattersall, Foster, & Johnston, Single-Layer Lookup Perceptions, IEE Proceedings-F, vol. 13, No. 3, pp. 46 54 Feb. 1991.
324Tattersall, Foster, & Johnston, Single-Layer Lookup Perceptions, IEE Proceedings-F, vol. 13, No. 3, pp. 46-54 Feb. 1991.
325Technical Visionary, Design News, pp. 74-86, Feb. 12, 1990.
326 *Teolis & Benedetto, Noise Suppression Using A Wavelet Model, IEEE 1994.
327TI Launches Second Generation Voice-Control PC Products, Design News, p. 44, Jun. 3, 1985.
328Tom & Tenorio, Short Utterance Recognition Using a Network with Minimum Training, Neural Networks vol. 4, pp. 711-722, 1991.
329 *Trancoso & Tribolet, Harmonic Postprocessing Speech Synthesized by Stochastic Coders, IEE Proceedings, vol. 136, Pt. 1, No. 2, pp. 141 144, Apr. 1989.
330Trancoso & Tribolet, Harmonic Postprocessing Speech Synthesized by Stochastic Coders, IEE Proceedings, vol. 136, Pt. 1, No. 2, pp. 141-144, Apr. 1989.
331Treumiet & Gong, Noise Independent Speech Recognition for a Variety of Noise Types, IEEE 1994.
332Tribolet, A New Phase Unwrapping Algorithm, IEEE Transactions on Acoustic, Speech and Signal Processing vol. ASSP-25, No. 2, pp. 170-177, Apr. 1977.
333Tunick, Signal-Processing Technique Takes Voice Coding to Extremes, Electronic Design, pp. 67-68, Aug. 6, 1987.
334 *Usagawa, Iwata & Ebata, Speech Parameter Extraction in Noisy Environment Using a Masking Model, IEEE 1994.
335 *Vaseghi, Milner & Humphries, Noisy Speech Recognition Using Cepstral-Time Features and Spectral-Tim Filters, IEEE 1994.
336Visser, Voice Recognition Fells Technical Barriers, CTM Technology, May 1987.
337Voice Recognizers Ignore Noise, Machine Design, p. 18, Dec. 12, 1985.
338Waibel & Hampshire, Building Blocks for Speech, BYTE, pp. 235-245, Aug. 1989.
339Wakita, Normalization of Vowels by Vocal-Tract Length and Its Application to Vowel Identification, IEE Transactions on Acoustics, Speech and Signal Processing, vol. ASSP-25, No. 2, pp. 183-192, Apr. 1977.
340Waldrop, A Landmark in Speech Recognition, Research News, p. 1615, Jun. 17, 1988.
341Wang, Liu, Lee & Chang, A Study on the Automatic Recognition of Voiceless Unaspirated Stops, J. Acoustical Society of America, vol. 89, No. 1, pp. 461-464, Jan. 1991.
342 *Wang, Wu, Chang & Lee, A Hierarchical Neural Network Model Based on C/V Segmentation Algorithm for Isolated Mandarin Speech Recognition, IEEE Transactions on Signal Processing, vol. 39, No. 9, pp. 2141 2147, Sep. 1991. Takahashi, Hamauchi, Tansho & Kimura, A Modularized Processor LSI with a Highly Parallel Structure for Continuous Speech Recognition, IEEE Journal of Solid-State Circuits, vol. 26, No. 6, pp. 833 843, Jun. 1991.
343Wang, Wu, Chang & Lee, A Hierarchical Neural Network Model Based on C/V Segmentation Algorithm for Isolated Mandarin Speech Recognition, IEEE Transactions on Signal Processing, vol. 39, No. 9, pp. 2141-2147, Sep. 1991. Takahashi, Hamauchi, Tansho & Kimura, A Modularized Processor LSI with a Highly Parallel Structure for Continuous Speech Recognition, IEEE Journal of Solid-State Circuits, vol. 26, No. 6, pp. 833-843, Jun. 1991.
344Watrous, Ladendorf & Kuhn, Complete Gradient Optimization of a Recurrent Network Applied to/b/,/d/,/g/Discrimination, J. Acoustical Society of America, vol. 87, No. 3, pp. 1301-1309, Mar. 1990.
345Wattenbarger, Garberg, Halpern & Lively, Serving Customers with Automatic Speech Recognition-Human Factors Issues, AT&T Technical Journal, pp. 28-41, May/Jun. 1993.
346Weitzman, Vowel Categorization and the Critical Band, Language and Speech, vol. 35, No, 1,2, pp. 115-125 1992.
347Wheddon & Linggard, A Novel Speech Noise Suppressor, Speech and Language Processing.
348Wheddon & Linggard, Speech and Language Processing, ISBN, 1990.
349Whipple, Low Residual Noise Speech Enhancement Utilizing Time-Frequency Filtering IEEE, 1994.
350White, Natural Language Understanding and Speech Recognition, Communications of the ACM, vol. 33, No. 8, pp. 72-82, Aug. 1990.
351Wightman, Shattuck-Hufnagel, Ostendorf & Price, Segmental Durations in the Vicinity of Prosopic Phrase Boundaries, J. Acoustical Society of America, vol. 91, No. 3, pp. 1707-1717, Mar. 1992.
352Wilkes & Cadzow, The Effects of Phase on High-Resolution Frequency Estimators, IEEE, Transactions on Signal Processing, vol. 41, No. 3, pp. 1319-1330, Mar. 1993.
353Wilpon, A Study on the Ability to Automatically Recognize Telephone-Quality Speech From Large Customer Populations, AT&T Technical Journal, vol. 64, No. 2, pp. 423-451, Feb. 1985.
354Wilpon, Mikkillneni, Roe & Gokcen, Speech Recognition: From the Laboratory to the Real World, AT& Technical Journal, pp. 14-23, Sep./Oct. 1990.
355Wu & Chan, Isolated Word Recognition by Neural Network Models With Cross-Correlation Coefficients for Speech Dynamics, IEEE Transactions on Pattern Analysis and Machine Intelligence vol. 15, No. 11, pp. 1174-1185, Nov. 1993.
356Wu & Childers, Gender Recognition From Speech. Part I: Coarse Analysis, J. Acoustical Society of America vol. 90, No. 4, pp. 1828-1856, Oct. 1991.
357 *Xie & Compernolle, A Family of MLP Based Nonlinear Spectral Estimators for Noise Reduction, IEEE 1994.
358 *Young Designing a Conversational Speech Interface, IEE Proceedings, vol. 133, Pt. E, No. 6, pp. 305 311, Nov. 1986.
359Young Designing a Conversational Speech Interface, IEE Proceedings, vol. 133, Pt. E, No. 6, pp. 305-311, Nov. 1986.
360 *Young, Competitive Training: A Connectionist Approach to the Discriminative Training of Hidden Markov Models, IEE Proceedings-1, vol. 1338, No. 1, pp. 61 68, Feb. 1991.
361Young, Competitive Training: A Connectionist Approach to the Discriminative Training of Hidden Markov Models, IEE Proceedings-1, vol. 1338, No. 1, pp. 61-68, Feb. 1991.
362Young, Hauptmann, Ward, Smith & Werner, High Level Knowledge Souces in Usable Speech Recognition Systems, Communications of the ACM, vol. 32, No. 2, pp. 183-194, Feb. 1989.
363Yuhas & Goldstein, Comparing Human and Neural network Lip Readers, J. Accoustical Society of America, pp. 598-600, Jul. 1991.
364 *Yuhas, Goldstein & Sejnowski, Integration of Acoustic and Visual Speech Signals Using Neural Networks, IEEE Communications Magazine, pp. 65 71, Nov. 1989.
365Yuhas, Goldstein & Sejnowski, Integration of Acoustic and Visual Speech Signals Using Neural Networks, IEEE Communications Magazine, pp. 65-71, Nov. 1989.
366 *Yuhas,Goldstein, Sejnowski & Jenkins, Neural Network Models of Sensory Integration for Improved Vowe Recognition, IEEE Proceedings, vol. 78, No. 10, pp. 1658 1668, Oct. 1990.
367Yuhas,Goldstein, Sejnowski & Jenkins, Neural Network Models of Sensory Integration for Improved Vowe Recognition, IEEE Proceedings, vol. 78, No. 10, pp. 1658-1668, Oct. 1990.
368Zahorian & Jafharghi, Speaker Normalization of Static and Dynamic Vowel Spectral Features, J. Acoustical Society of America, vol. 90, No. 1, pp. 67-75, Jul. 1991.
369Zera, Onsan, Nguyen, & Green Auditory Profile Analysis of Harmonic Signals, J. Acoustical Society of America, vol. 93, No. 6, pp. 3431-3441, Jun. 1993.
370 *Zhang, Alder & Togneri, Using Gaussian Mixture Modeling in Speech Recognition, IEEE 1994.
371 *Zhao, Atlas & Zhuang, Application of the Gibbs Distribution to Hidden Markov Modeling in Speaker Indepdent Isolated Word Recognition, IEEE Transactions on Signal Processing, vol. 39, No. 6, pp. 1291 1299, Jun. 1991.
372Zhao, Atlas & Zhuang, Application of the Gibbs Distribution to Hidden Markov Modeling in Speaker Indepdent Isolated Word Recognition, IEEE Transactions on Signal Processing, vol. 39, No. 6, pp. 1291-1299, Jun. 1991.
373Zollo, Digital Filter Handles 24-Bit Data, Electronics Week, pp. 105-106, Oct. 22, 1984.
374Zue, Automatic Speech Recognition and Understanding, MIT Survey, pp. 185-200.
375 *Zue, The Use of Speech Knowledge in Automatic Speech Recognition, lEE Proceedings, Vol. 73, No. 11, pp. 1602 1615, Nov. 1985.
376Zue, The Use of Speech Knowledge in Automatic Speech Recognition, lEE Proceedings, Vol. 73, No. 11, pp. 1602-1615, Nov. 1985.
Referenced by
Citing PatentFiling datePublication dateApplicantTitle
US5797116 *Jan 21, 1997Aug 18, 1998Canon Kabushiki KaishaVoice communication method
US5839099 *Jun 11, 1996Nov 17, 1998Guvolt, Inc.Signal conditioning apparatus
US5873062 *Jan 9, 1997Feb 16, 1999Fonix CorporationUser independent, real-time speech recognition system and method
US5899974 *Dec 31, 1996May 4, 1999Intel CorporationCompressing speech into a digital format
US5909666 *Jun 26, 1997Jun 1, 1999Dragon Systems, Inc.Speech recognition system which creates acoustic models by concatenating acoustic models of individual words
US5915236 *Jun 26, 1997Jun 22, 1999Dragon Systems, Inc.Word recognition system which alters code executed as a function of available computational resources
US5920836 *Jun 26, 1997Jul 6, 1999Dragon Systems, Inc.Word recognition system using language context at current cursor position to affect recognition probabilities
US5920837 *Jun 26, 1997Jul 6, 1999Dragon Systems, Inc.Word recognition system which stores two models for some words and allows selective deletion of one such model
US5983179 *Jun 26, 1997Nov 9, 1999Dragon Systems, Inc.Speech recognition system which turns its voice response on for confirmation when it has been turned off without confirmation
US6061654 *Dec 16, 1996May 9, 2000At&T Corp.System and method of recognizing letters and numbers by either speech or touch tone recognition utilizing constrained confusion matrices
US6073097 *Jun 26, 1997Jun 6, 2000Dragon Systems, Inc.Speech recognition system which selects one of a plurality of vocabulary models
US6092043 *Jun 26, 1997Jul 18, 2000Dragon Systems, Inc.Apparatuses and method for training and operating speech recognition systems
US6101468 *Jun 26, 1997Aug 8, 2000Dragon Systems, Inc.Apparatuses and methods for training and operating speech recognition systems
US6122612 *Nov 20, 1997Sep 19, 2000At&T CorpCheck-sum based method and apparatus for performing speech recognition
US6122615 *Mar 24, 1998Sep 19, 2000Fujitsu LimitedSpeech recognizer using speaker categorization for automatic reevaluation of previously-recognized speech data
US6137863 *Dec 13, 1996Oct 24, 2000At&T Corp.Statistical database correction of alphanumeric account numbers for speech recognition and touch-tone recognition
US6141661 *Oct 17, 1997Oct 31, 2000At&T CorpMethod and apparatus for performing a grammar-pruning operation
US6154579 *Aug 11, 1997Nov 28, 2000At&T Corp.Confusion matrix based method and system for correcting misrecognized words appearing in documents generated by an optical character recognition technique
US6170049 *Aug 11, 1999Jan 2, 2001Texas Instruments IncorporatedPC circuits, systems and methods
US6205261Feb 5, 1998Mar 20, 2001At&T Corp.Confusion set based method and system for correcting misrecognized words appearing in documents generated by an optical character recognition technique
US6205428Nov 20, 1997Mar 20, 2001At&T Corp.Confusion set-base method and apparatus for pruning a predetermined arrangement of indexed identifiers
US6223158Feb 4, 1998Apr 24, 2001At&T CorporationStatistical option generator for alpha-numeric pre-database speech recognition correction
US6224384 *Jun 27, 2000May 1, 2001Scientific Learning Corp.Method and apparatus for training of auditory/visual discrimination using target and distractor phonemes/graphemes
US6290504 *Oct 8, 1999Sep 18, 2001Scientific Learning Corp.Method and apparatus for reporting progress of a subject using audio/visual adaptive training stimulii
US6328569 *Jun 26, 1998Dec 11, 2001Scientific Learning Corp.Method for training of auditory/visual discrimination using target and foil phonemes/graphemes within an animated story
US6331115 *Jun 30, 1998Dec 18, 2001Scientific Learning Corp.Method for adaptive training of short term memory and auditory/visual discrimination within a computer game
US6334776 *Jun 27, 2000Jan 1, 2002Scientific Learning CorporationMethod and apparatus for training of auditory/visual discrimination using target and distractor phonemes/graphemes
US6400805Jun 15, 1998Jun 4, 2002At&T Corp.Statistical database correction of alphanumeric identifiers for speech recognition and touch-tone recognition
US6487531Jul 6, 1999Nov 26, 2002Carol A. TosayaSignal injection coupling into the human vocal tract for robust audible and inaudible voice recognition
US6507819 *Feb 18, 2000Jan 14, 2003Matsushita Electric Industrial Co., Ltd.Sound signal processor for extracting sound signals from a composite digital sound signal
US6599129Sep 24, 2001Jul 29, 2003Scientific Learning CorporationMethod for adaptive training of short term memory and auditory/visual discrimination within a computer game
US6931292Jun 19, 2000Aug 16, 2005Jabra CorporationNoise reduction method and apparatus
US6990445Dec 17, 2001Jan 24, 2006Xl8 Systems, Inc.System and method for speech recognition and transcription
US7082395Nov 27, 2002Jul 25, 2006Tosaya Carol ASignal injection coupling into the human vocal tract for robust audible and inaudible voice recognition
US7130705Jan 8, 2001Oct 31, 2006International Business Machines CorporationSystem and method for microphone gain adjust based on speaker orientation
US7162426 *Oct 2, 2000Jan 9, 2007Xybernaut CorporationComputer motherboard architecture with integrated DSP for continuous and command and control speech processing
US7319959 *May 14, 2003Jan 15, 2008Audience, Inc.Multi-source phoneme classification for noise-robust automatic speech recognition
US7324984 *Nov 14, 2003Jan 29, 2008Intel CorporationMethod and system for detecting semantic events
US7392178 *Feb 26, 2003Jun 24, 2008Electronic Navigation Research Institute, An Independent Administration InstitutionChaos theoretical diagnosis sensitizer
US7401017Apr 4, 2006Jul 15, 2008Nuance CommunicationsAdaptive multi-pass speech recognition system
US7454345 *Feb 23, 2005Nov 18, 2008Fujitsu LimitedWord or collocation emphasizing voice synthesizer
US7519531 *Mar 30, 2005Apr 14, 2009Microsoft CorporationSpeaker adaptive learning of resonance targets in a hidden trajectory model of speech coarticulation
US7523038 *Jul 24, 2003Apr 21, 2009Arie AriavVoice controlled system and method
US7555430 *Apr 4, 2006Jun 30, 2009Nuance CommunicationsSelective multi-pass speech recognition system and method
US7565284Mar 1, 2005Jul 21, 2009Microsoft CorporationAcoustic models with structured hidden dynamics with integration over many possible hidden trajectories
US7565292Sep 17, 2004Jul 21, 2009Micriosoft CorporationQuantitative model for formant dynamics and contextually assimilated reduction in fluent speech
US7630899Sep 7, 2006Dec 8, 2009At&T Intellectual Property Ii, L.P.Concise dynamic grammars using N-best selection
US7646293 *Jun 5, 2006Jan 12, 2010Hong Fu Jin Precision Industry (Shen Zhen) Co., Ltd.System and method for testing a buzzer associated with a computer
US7813925 *Apr 6, 2006Oct 12, 2010Canon Kabushiki KaishaState output probability calculating method and apparatus for mixture distribution HMM
US7890325 *Mar 16, 2006Feb 15, 2011Microsoft CorporationSubword unit posterior probability for measuring confidence
US7895065Dec 17, 2003Feb 22, 2011Sony CorporationMethod and apparatus for an itinerary planner
US7937260Jun 15, 1998May 3, 2011At&T Intellectual Property Ii, L.P.Concise dynamic grammars using N-best selection
US7966177 *Aug 13, 2001Jun 21, 2011Hans GeigerMethod and device for recognising a phonetic sound sequence or character sequence
US8050948Jan 11, 2011Nov 1, 2011Sony CorporationMethod and apparatus for an itinerary planner
US8050949Jan 11, 2011Nov 1, 2011Sony CorporationMethod and apparatus for an itinerary planner
US8090077Apr 2, 2007Jan 3, 2012Microsoft CorporationTesting acoustic echo cancellation and interference in VoIP telephones
US8099290 *Oct 20, 2009Jan 17, 2012Mitsubishi Electric CorporationVoice recognition device
US8364492 *Jul 6, 2007Jan 29, 2013Nec CorporationApparatus, method and program for giving warning in connection with inputting of unvoiced speech
US8447605 *Jun 3, 2005May 21, 2013Nintendo Co., Ltd.Input voice command recognition processing apparatus
US8542802Feb 15, 2007Sep 24, 2013Global Tel*Link CorporationSystem and method for three-way call detection
US8599704Jan 23, 2007Dec 3, 2013Microsoft CorporationAssessing gateway quality using audio systems
US8630726Feb 12, 2009Jan 14, 2014Value-Added Communications, Inc.System and method for detecting three-way call circumvention attempts
US8682665Apr 28, 2011Mar 25, 2014At&T Intellectual Property Ii, L.P.Concise dynamic grammars using N-best selection
US8731934Feb 15, 2008May 20, 2014Dsi-Iti, LlcSystem and method for multi-modal audio mining of telephone conversations
US8737573 *May 9, 2012May 27, 2014Intelligent Decisions, Inc.Systems, methods, and devices for testing communication lines
US20090254350 *Jul 6, 2007Oct 8, 2009Nec CorporationApparatus, Method and Program for Giving Warning in Connection with inputting of unvoiced Speech
US20100042237 *Jul 24, 2009Feb 18, 2010Chi Mei Communication Systems, Inc.Mobile communication device and audio signal adjusting method thereof
US20130028452 *Feb 6, 2012Jan 31, 2013Makoto NishizakiHearing aid adjustment device
US20130046805 *Nov 11, 2010Feb 21, 2013Paul Reed Smith Guitars Limited PartnershipPrecision Measurement of Waveforms Using Deconvolution and Windowing
US20130121479 *May 9, 2012May 16, 2013Intelligent Decisions, Inc.Systems, methods, and devices for testing communication lines
WO2000016312A1 *Sep 2, 1999Mar 23, 2000Sony Electronics IncMethod for implementing a speech verification system for use in a noisy environment
Classifications
U.S. Classification704/254, 704/E15.015, 704/270.1, 704/E15.004
International ClassificationG10L25/93, G10L15/02, G10L15/10
Cooperative ClassificationG10L25/09, G10L25/93, G10L15/02, G10L15/10, G10L25/18
European ClassificationG10L15/02, G10L15/10
Legal Events
DateCodeEventDescription
Aug 4, 2009FPExpired due to failure to pay maintenance fee
Effective date: 20090617
Jun 17, 2009LAPSLapse for failure to pay maintenance fees
Dec 22, 2008REMIMaintenance fee reminder mailed
Mar 31, 2005SULPSurcharge for late payment
Year of fee payment: 7
Mar 31, 2005FPAYFee payment
Year of fee payment: 8
Jan 5, 2005REMIMaintenance fee reminder mailed
Apr 4, 2001SULPSurcharge for late payment
Apr 4, 2001FPAYFee payment
Year of fee payment: 4
Jan 9, 2001REMIMaintenance fee reminder mailed
Apr 13, 1999CCCertificate of correction
Jan 20, 1995ASAssignment
Owner name: FONIX CORPORATION, UTAH
Free format text: ASSIGNMENT OF ASSIGNORS INTEREST;ASSIGNORS:HANSEN, C. HAL;SHEPARD, DALE LYNN;MONCUR, ROBERT BRIAN;REEL/FRAME:007328/0254
Effective date: 19950113